Welcome to Elliot’s documentation!¶

Elliot is a comprehensive recommendation framework that analyzes the recommendation problem from the researcher’s perspective. It conducts a whole experiment, from dataset loading to results gathering. The core idea is to feed the system with a simple and straightforward configuration file that drives the framework through the experimental setting choices. Elliot untangles the complexity of combining splitting strategies, hyperparameter model optimization, model training, and the generation of reports of the experimental results.

system schema¶

The framework loads, filters, and splits the data considering a vast set of strategies (splitting methods and filtering approaches, from temporal training-test splitting to nested K-folds Cross-Validation). Elliot optimizes hyperparameters for several recommendation algorithms, selects the best models, compares them with the baselines providing intra-model statistics, computes metrics spanning from accuracy to beyond-accuracy, bias, and fairness, and conducts statistical analysis (Wilcoxon and Paired t-test).

Elliot aims to keep the entire experiment reproducible and put the user in control of the framework.

Introduction¶

Elliot is a comprehensive recommendation framework that analyzes the recommendation problem from the researcher’s perspective. It conducts a whole experiment, from dataset loading to results gathering. The core idea is to feed the system with a simple and straightforward configuration file that drives the framework through the experimental setting choices. Elliot untangles the complexity of combining splitting strategies, hyperparameter model optimization, model training, and the generation of reports of the experimental results.

system schema¶

The framework loads, filters, and splits the data considering a vast set of strategies (splitting methods and filtering approaches, from temporal training-test splitting to nested K-folds Cross-Validation). Elliot optimizes hyperparameters for several recommendation algorithms, selects the best models, compares them with the baselines providing intra-model statistics, computes metrics spanning from accuracy to beyond-accuracy, bias, and fairness, and conducts statistical analysis (Wilcoxon and Paired t-test).

Elliot aims to keep the entire experiment reproducible and put the user in control of the framework.

For all the details about Elliot, please refer to the paper and cite [Elliot]

Elliot: Vito Walter Anelli and Alejandro Bellogín and Antonio Ferrara and Daniele Malitesta and Felice Antonio Merra and Claudio Pomo and Francesco Maria Donini and Tommaso Di Noia. 2021. Elliot: a Comprehensive and Rigorous Framework for Reproducible Recommender Systems Evaluation. Under review. arXiv:2103.02590 [cs.IR].

Install Elliot¶

Elliot works with the following operating systems:

Linux
Windows 10
macOS X

Elliot requires Python version 3.6 or later.

Elliot requires tensorflow version 2.3.2 or later. If you want to use Elliot with GPU, please ensure that CUDA or cudatoolkit version is 7.6 or later. This requires NVIDIA driver version >= 10.1 (for Linux and Windows10).

Please refer to this document for further working configurations.

Install from source¶

CONDA¶

git clone https://github.com//sisinflab/elliot.git && cd elliot
conda create --name elliot_env python=3.8
conda activate
pip install --upgrade pip
pip install -e . --verbose

VIRTUALENV¶

git clone https://github.com//sisinflab/elliot.git && cd elliot
virtualenv -p /usr/bin/pyhton3.6 venv # your python location and version
source venv/bin/activate
pip install --upgrade pip
pip install -e . --verbose

Quick Start¶

Hello Word Cofiguration¶

Elliot’s entry point is the function run_experiment, which accepts a configuration file that drives the whole experiment. In the following, a sample configuration file is shown to demonstrate how a sample and explicit structure can generate a rigorous experiment.

from elliot.run import run_experiment

run_experiment("configuration/file/path")

The following file is a simple configuration for an experimental setup. It contains all the instructions to get the MovieLens-1M catalog from a specific path and perform a train test split in a random sample way with a ratio of 20%.

This experiment provides a hyperparameter optimization with a grid search strategy for an Item-KNN model. Indeed, it is seen that the possible values of neighbors are closed in squared brackets. It indicates that two different models equipped with two different neighbors’ values will be trained and compared to select the best configuration. Moreover, this configuration obliges Elliot to save the recommendation lists with at most 10 items per user as suggest by top_k property.

In this basic experiment, only a simple metric is considered in the final evaluation study. The candidate metric is nDCG for a cutoff equal to top_k, unless otherwise noted.

experiment:
  dataset: movielens_1m
  data_config:
    strategy: dataset
    dataset_path: ../data/movielens_1m/dataset.tsv
    splitting:
      test_splitting:
        strategy: random_subsampling
        test_ratio: 0.2
    models:
      ItemKNN:
        meta:
          hyper_opt_alg: grid
          save_recs: True
        neighbors: [50, 100]
        similarity: cosine
    evaluation:
      simple_metrics: [nDCG]
    top_k: 10

Basic Configuration¶

In the first scenario, the experiments require comparing a group of RSs whose parameters are optimized via a grid-search.

The configuration specifies the data loading information, i.e., semantic features source files, in addition to the filtering and splitting strategies.

In particular, the latter supplies an entirely automated way of preprocessing the dataset, which is often a time-consuming and non-easily-reproducible phase.

The simple_metrics field allows computing accuracy and beyond-accuracy metrics, with two top-k cut-off values (5 and 10) by merely inserting the list of desired measures, e.g., [Precision, nDCG, …]. The knowledge-aware recommendation model, AttributeItemKNN, is compared against two baselines: Random and ItemKNN, along with a user-implemented model that is external.MostPop.

The configuration makes use of elliot’s feature of conducting a grid search-based hyperparameter optimization strategy by merely passing a list of possible hyperparameter values, e.g., neighbors: [50, 70, 100].

The reported models are selected according to nDCG@10.

To see the full configuration file please visit the following link_basic

To run the experiment use the following script_basic

Advanced Configuration¶

The second scenario depicts a more complex experimental setting. In the configuration, the user specifies an elaborate data splitting strategy, i.e., random_subsampling (for test splitting) and random_cross_validation (for model selection), by setting few splitting configuration fields.

The configuration does not provide a cut-off value, and thus a top-k field value of 50 is assumed as the cut-off.

Moreover, the evaluation section includes the UserMADrating metric.

Elliot considers it as a complex metric since it requires additional arguments.

The user also wants to implement a more advanced hyperparameter tuning optimization. For instance, regarding NeuMF, Bayesian optimization using Tree of Parzen Estimators is required (i.e., hyper_opt_alg: tpe) with a logarithmic uniform sampling for the learning rate search space.

Moreover, Elliot allows considering complex neural architecture search spaces by inserting lists of tuples. For instance, (32, 16, 8) indicates that the neural network consists of three hidden layers with 32, 16, and 8 units, respectively.

To see the full configuration file please visit the following link_advanced

To run the experiment use the following script_advanced

Configuration file¶

Input Data Configuration¶

The first key component of the config file is the data_config section.

experiment:
  data_config:
    strategy: dataset|fixed|hierarchy
    dataloader: KnowledgeChainsLoader|DataSetLoader
    dataset_path: this/is/the/path.tsv
    root_folder: this/is/the/path
    train_path: this/is/the/path.tsv
    validation_path: this/is/the/path.tsv
    test_path: this/is/the/path.tsv
    side_information:
        feature_data: this/is/the/path.tsv
        map: this/is/the/path.tsv
        features: this/is/the/path.tsv
        properties: this/is/the/path.conf

In this section, we can define which input files and how they should be loaded.

In the following, we will consider as datasets, tab-separated-value files that contain one interaction per row, in the format:

UserID ItemID Rating [ TimeStamp ]

where TimeStamp is optional.

Strategies¶

According to the kind of data we have, we can choose among three different loading strategies: dataset, fixed, hierarchy.

dataset assumes that the input data is NOT previously split in training, validation, and test set. For this reason, ONLY if we adopt a dataset strategy we can later perform prefiltering and splitting operations.

dataset takes just ONE default parameter: dataset_path, which points to the stored dataset.

experiment:
  data_config:
    strategy: dataset
    dataset_path: this/is/the/path.tsv

fixed strategy assumes that our data has been previously split into training/validation/test sets or training/test sets. Since data is supposed as previously split, no further prefiltering and splitting operation is contemplated.

fixed takes two mandatory parameters: train_path and test_path, and one optional parameter, validation_path.

experiment:
  data_config:
    strategy: fixed
    train_path: this/is/the/path.tsv
    validation_path: this/is/the/path.tsv
    test_path: this/is/the/path.tsv

The last strategy is hierarchy. hierarchy is designed to load a dataset that has been previously split and filtered with Elliot. Here, the data is assumed as split and no further prefiltering and splitting operations are needed.

hierarchy takes one mandatory parameter, root_folder, that points to the folder where we previously stored the split files.

experiment:
  data_config:
    strategy: hierarchy
    root_folder: this/is/the/path

Data Loaders¶

Within the data_config section, we can also enable data-specific Data Loaders. Each Data Loader is designed to handle a specific kind of additional data.

It is possible to enable a Data Loader by inserting the field dataloader and passing the corresponding name. For instance, the Visual Data Loader lets the user consider precomputed visual feature vectors or (inclusive) images.

To pass the required parameters to the Data Loader, we use a specific subsection, named side_information. There we can enable the required (by the specific Data Loader) fields and insert the corresponding values.

An example can be:

experiment:
  data_config:
    strategy: fixed
    dataloader: VisualLoader
    train_path: this/is/the/path.tsv
    test_path: this/is/the/path.tsv
    side_information:
        feature_data: this/is/the/path/to/features.npy

For further details regarding the Data Loaders, please refer to the section.

Data Prefiltering¶

Elliot provides several prefiltering strategies. To enable Prefiltering operations, we can insert the corresponding block into our config file:

experiment:
  prefiltering:
    strategy: global_threshold|user_average|user_k_core|item_k_core|iterative_k_core|n_rounds_k_core|cold_users
    threshold: 3|average
    core: 5
    rounds: 2

In detail, Elliot provides eight main prefiltering approaches: global_threshold, user_average, user_k_core, item_k_core, iterative_k_core, n_rounds_k_core, cold_users.

global_threshold assumes a single system-wise threshold to filter out irrelevant transactions. global_threshold takes one mandatory parameter, threshold. threshold takes, as values, a float (ratings >= threshold will be kept), or the string average. With average, the system computes the global mean of the rating values and filters out all the ratings below.

experiment:
  prefiltering:
    strategy: global_threshold
    threshold: 3

experiment:
  prefiltering:
    strategy: global_threshold
    threshold: average

user_average has no parameters, and the system filters out the ratings below each user rating values mean.

experiment:
  prefiltering:
    strategy: user_average

user_k_core filters out all the users with a number of transactions lower than the given k core. It takes a parameter, core, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: user_k_core
    core: 5

item_k_core filters out all the items with a number of transactions lower than the given k core. It takes a parameter, core, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: item_k_core
    core: 5

iterative_k_core runs iteratively user_k_core, and item_k_core until the dataset is no further modified. It takes a parameter, core, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: iterative_k_core
    core: 5

n_rounds_k_core runs iteratively user_k_core, and item_k_core for a specified number of rounds. It takes the first parameter, core, where the user passes an int corresponding to the desired value. It takes the second parameter, rounds, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: n_rounds_k_core
    core: 5
    rounds: 2

cold_users filters out all the users with a number of interactions higher than a given threshold. It takes a parameter, threshold, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: cold_users
    threshold: 3

Data Splitting¶

Elliot provides several splitting strategies. To enable the splitting operations, we can insert the corresponding section:

experiment:
  splitting:
    save_on_disk: True
    save_folder: this/is/the/path/
    test_splitting:
        strategy: fixed_timestamp|temporal_hold_out|random_subsampling|random_cross_validation
        timestamp: best|1609786061
        test_ratio: 0.2
        leave_n_out: 1
        folds: 5
    validation_splitting:
        strategy: fixed_timestamp|temporal_hold_out|random_subsampling|random_cross_validation
        timestamp: best|1609786061
        test_ratio: 0.2
        leave_n_out: 1
        folds: 5

Before deepening the splitting configurations, we can configure Elliot to save on disk the split files, once the splitting operation is completed.

To this extent, we can insert two fields into the section: save_on_disk, and save_folder.

save_on_disk enables the writing process, and save_folder specifies the system location where to save the split files:

experiment:
  splitting:
    save_on_disk: True
    save_folder: this/is/the/path/

Now, we can insert one (or two) specific subsections to detail the train/test, and the train/validation splitting via the corresponding fields: test_splitting, and validation_splitting. test_splitting is clearly mandatory, while validation_splitting is optional. Since the two subsections follow the same guidelines, here we detail test_splitting without loss of generality.

Elliot enables four splitting families: fixed_timestamp, temporal_hold_out, random_subsampling, random_cross_validation.

fixed_timestamp assumes that there will be a specific timestamp to split prior interactions (train) and future interactions. It takes the parameter timestamp, that can assume one of two possible kind of values: a long corresponding to a specific timestamp, or the string best computed following Anelli et al.

experiment:
  splitting:
    test_splitting:
        strategy: fixed_timestamp
        timestamp: 1609786061

experiment:
  splitting:
    test_splitting:
        strategy: fixed_timestamp
        timestamp: best

temporal_hold_out relies on a temporal split of user transactions. The split can be realized following two different approaches: a ratio-based and a leave-n-out-based approach. If we enable the test_ratio field with a float value, Elliot splits data retaining the last (100 * test_ratio) % of the user transactions for the test set. If we enable the leave_n_out field with an int value, Elliot retains the last leave_n_out transactions for the test set.

experiment:
  splitting:
    test_splitting:
        strategy: temporal_hold_out
        test_ratio: 0.2

experiment:
  splitting:
    test_splitting:
        strategy: temporal_hold_out
        leave_n_out: 1

random_subsampling generalizes random hold-out strategy. It takes a test_ratio parameter with a float value to define the train/test ratio for user-based hold-out splitting. Alternatively, it can take leave_n_out with an int value to define the number of transaction retained for the test set. Moreover, the splitting operation can be repeated enabling the folds field and passing an int. In that case, the overall splitting strategy corresponds to a user-based random subsampling strategy.

experiment:
  splitting:
    test_splitting:
        strategy: random_subsampling
        test_ratio: 0.2

experiment:
  splitting:
    test_splitting:
        strategy: random_subsampling
        test_ratio: 0.2
        folds: 5

experiment:
  splitting:
    test_splitting:
        strategy: random_subsampling
        leave_n_out: 1
        folds: 5

random_cross_validation adopts a k-folds cross-validation splitting strategy. It takes the parameter folds with an int value, that defines the overall number of folds to consider.

experiment:
  splitting:
    test_splitting:
        strategy: random_cross_validation
        folds: 5

Dataset Name Configuration¶

Elliot needs a MANDATORY field, dataset, that identifies the name of the dataset used for the experiment. This information is used in the majority of the experimental steps, to identify the experiment and save the files correctly:

experiment:
  dataset: dataset_name

Output Configuration¶

Elliot lets the user specify where to store specific output files: the recommendation lists, the model weights, the evaluation results, and the logs:

experiment:
  path_output_rec_result: this/is/the/path/
  path_output_rec_weight: this/is/the/path/
  path_output_rec_performance: this/is/the/path/
  path_log_folder: this/is/the/path/

path_output_rec_result lets the user define the path to the folder to store the recommendation lists.

path_output_rec_weight lets the user define the path to the folder to store the model weights.

path_output_rec_performance lets the user define the path to the folder to store the evaluation results.

path_log_folder lets the user define the path to the folder to store the logs.

If not provided, Elliot creates a results folder in the parent folder of the config file location.

Inside it, Elliot creates an experiment-specific folder with the name of the dataset, and there it creates the recs/, weights/, and performance/ folders, respectively.

Moreover, Elliot creates a log/ folder in the parent folder of the config file location.

Evaluation Configuration¶

Elliot provides several facilities to evaluate recommendation systems. The majority of the evaluation techniques require the computation of user-specific recommendation lists (some techniques use recommendation systems to perform knowledge completion or other tasks).

To define the length of the user recommendation list, Elliot provides a specific mandatory field, top_k, that takes an int representing the list length.

Beyond the former general definition, to specify the evaluation configuration, we can insert a specific section:

experiment:
  top_k: 50
  evaluation:
    cutoffs: [10, 5]
    simple_metrics: [ nDCG, Precision, Recall]
    relevance_threshold: 1
    paired_ttest: True
    wilcoxon_test: True
    complex_metrics:
    - metric: DSC
      beta: 2
    - metric: SRecall
      feature_data: this/is/the/path.tsv

In that section, we can detail the main characteristics of our experimental benchmark.

In particular, we can provide Elliot with the information regarding the metrics we want to compute. According to the metrics definition, some of them might require additional parameters or files. To make it easier for the user to pass metrics and optional arguments, Elliot partitions the metrics into simple_metrics and complex_metrics.

simple_metrics can be inserted as a field into the evaluation section, and it takes as a value the list of the metrics we want to compute. In the simple metrics set, we find all the metrics that DO NOT require any other additional parameter or file:

experiment:
  top_k: 50
  evaluation:
    cutoffs: [10, 5]
    simple_metrics: [ nDCG, Precision, Recall]
    relevance_threshold: 1

The majority of the evaluation metrics relies on the notions of cut-off and relevance threshold.

The cut-off is the maximum length of the recommendation list we want to consider when computing the metric (it could be different from the top k). To pass cut-off values, we can enable the cutoffs field and pass a single value or a list of values. Elliot will compute the evaluation results for each considered cut-off. If cutoffs field is not provided, top_k value is assumed as a cut-off.

The relevance threshold is the minimum value of the rating to consider a test transaction relevant for the evaluation process. We can pass the relevance threshold value to the corresponding relevance_threshold field. If not given, relevance_threshold is set to 0.

The set of metrics that require additional arguments is referred to as complex_metrics. The inclusion of the metrics follows the syntax:

experiment:
  evaluation:
    complex_metrics:
    - metric: complex_metric_name_0
      parameter_0: 2
    - metric: complex_metric_name_1
      parameter_1: this/is/the/path.tsv

where parameter_0 and parameter_1 are metric-specific parameters of any kind.

For further details about the available metrics, please see the corresponding section.

Finally, Elliot enables the computation of paired statistical hypothesis tests, namely, Wilcoxon, and Student’s paired t-tests.

To enable them, we can insert the corresponding boolean fields into the evaluation section:

experiment:
  evaluation:
    paired_ttest: True
    wilcoxon_test: True

All the evaluation results are available in the performance folder at the end of the experiment.

Print evaluation results as triples¶

It is common in the Recommender Systems community to generate the evaluation tables with the format: [method,metric,value].

This choice easily lets use custom pivot tables on the data, and thus enabling several complex analysis. To obtain additional evaluation summaries in this format, insert the following field:

experiment:
  print_results_as_triplets: True

Test the config file¶

Since an experiment may take a long time, a possible error in the configuration file in the last model configuration can lead to a severe waste of time. To avoid common mistakes in config file creation, Elliot provides a specific field that tests our configuration file before the actual run of the experiment. The feature can be activated as follows:

experiment:
  config_test: True

GPU Acceleration¶

Elliot lets the user enable GPU acceleration with Tensorflow. To select the gpu on which we can run our experiments, use the following syntax:

experiment:
  gpu: 1

If a negative value is passed, or the field is missing, the computation will take place on the CPU.

Please note that the configuration of tensorflow to work with GPUs is not covered by this guide. Please refer to the Tensorflow documentation for that.

Recommendation Model Configuration¶

To include the recommendation models, Elliot provides a straightforward syntax.

First, we can create a new section in the experiment, named models:

experiment:
  models:

Then, we can insert a list of recommendation models in which each model respects the following syntax:

experiment:
  models:
    model_0:
      meta:
        meta_parameter_0: something
      model_parameter_0: something
      model_parameter_1: something
      model_parameter_2: something
    model_1:
      meta:
        meta_parameter_0: something
      model_parameter_0: something
      model_parameter_1: something
      model_parameter_2: something

meta is a mandatory field that lets the user define some parameters that all recommendation models share, but they can decline differently.

The decision to save model weights and recommendations, the choice of the validation metric and cut-off, the chosen hyperparameter tuning strategy, the verbosity, and the frequency of the evaluation during the training belong to this category.

In detail, use:

verbose boolean field to enable verbose logs

save_recs boolean field to enable recommendation lists storage

save_weights boolean field to enable model weights storage

validation_metric mixed field (string @ int) to define the simple metric and the cut-off used for the model selection. If not provided it takes the first provided simple metric, and the first cut-off.

validation_rate int field: where applicable, define the iteration interval for the validation and test evaluation

hyper_opt_alg string field: it defines the hyperparameter tuning strategy

hyper_max_evals int field: where applicable, it defines the number of samples to consider for hyperparameter evaluation

To fully understand how to conduct hyperparameter optimization in Elliot, please refer to the corresponding section.

Finally, model_parameter_0, model_parameter_1, and model_parameter_2 represents the model-specific parameters.

For further details on model-specific parameters see the corresponding section.

Example:

experiment:
  models:
    KaHFMEmbeddings:
      meta:
        hyper_max_evals: 20
        hyper_opt_alg: tpe
        validation_rate: 1
        verbose: True
        save_weights: True
        save_recs: True
        validation_metric: nDCG@10
      epochs: 100
      batch_size: -1
      lr: 0.0001
      l_w: 0.005
      l_b: 0

Data Preparation¶

The first key component of the config file is the data_config section.

experiment:
  data_config:
    strategy: dataset|fixed|hierarchy
    dataloader: KnowledgeChainsLoader|DataSetLoader
    dataset_path: this/is/the/path.tsv
    root_folder: this/is/the/path
    train_path: this/is/the/path.tsv
    validation_path: this/is/the/path.tsv
    test_path: this/is/the/path.tsv
    side_information:
        feature_data: this/is/the/path.tsv
        map: this/is/the/path.tsv
        features: this/is/the/path.tsv
        properties: this/is/the/path.conf

In this section, we can define which input files and how they should be loaded.

In the following, we will consider as datasets, tab-separated-value files that contain one interaction per row, in the format:

UserID ItemID Rating [ TimeStamp ]

where TimeStamp is optional.

Data preparation Strategies¶

According to the kind of data we have, we can choose among three different loading strategies: dataset, fixed, hierarchy.

dataset assumes that the input data is NOT previously split in training, validation, and test set. For this reason, ONLY if we adopt a dataset strategy we can later perform prefiltering and splitting operations.

dataset takes just ONE default parameter: dataset_path, which points to the stored dataset.

experiment:
  data_config:
    strategy: dataset
    dataset_path: this/is/the/path.tsv

fixed strategy assumes that our data has been previously split into training/validation/test sets or training/test sets. Since data is supposed as previously split, no further prefiltering and splitting operation is contemplated.

fixed takes two mandatory parameters: train_path and test_path, and one optional parameter, validation_path.

experiment:
  data_config:
    strategy: fixed
    train_path: this/is/the/path.tsv
    validation_path: this/is/the/path.tsv
    test_path: this/is/the/path.tsv

The last strategy is hierarchy. hierarchy is designed to load a dataset that has been previously split and filtered with Elliot. Here, the data is assumed as split and no further prefiltering and splitting operations are needed.

hierarchy takes one mandatory parameter, root_folder, that points to the folder where we previously stored the split files.

experiment:
  data_config:
    strategy: hierarchy
    root_folder: this/is/the/path

Loading Data¶

RSs experiments could require different data sources such as user-item feedback or additional side information,e.g., the visual features of an item images. To fulfill these requirements, Elliot comes with different implementations of the Loading module. Additionally, the user can design computationally expensive prefiltering and splitting procedures that can be stored and loaded to save future computation. Data-driven extensions can handle additional data like visual features, and semantic features extracted from knowledge graphs. Once a side-information-aware Loading module is chosen, it filters out the items devoiding the required information to grant a fair comparison.

Within the data_config section, we can also enable data-specific Data Loaders. Each Data Loader is designed to handle a specific kind of additional data.

It is possible to enable a Data Loader by inserting the field dataloader and passing the corresponding name. For instance, the Visual Data Loader lets the user consider precomputed visual feature vectors or (inclusive) images.

To pass the required parameters to the Data Loader, we use a specific subsection, named side_information. There we can enable the required (by the specific Data Loader) fields and insert the corresponding values.

An example can be:

experiment:
  data_config:
    strategy: fixed
    dataloader: VisualLoader
    train_path: this/is/the/path.tsv
    test_path: this/is/the/path.tsv
    side_information:
        feature_data: this/is/the/path/to/features.npy

Data Loaders¶

Work in progress

Prefiltering Data¶

After data loading,Elliot provides data filtering operations through two possible strategies. The first strategy implemented in the Prefiltering module isFilter-by-rating, which drops off a user-item interaction if the preference score is smaller than a given threshold. It can be (i) a Numerical value, e.g.,3.5, (ii)a Distributional detail, e.g., global rating average value, or (iii) a user-based distributional (User Dist.) value, e.g., user’s average rat-ing value. The second prefiltering strategy,𝑘-core, filters out users,items, or both, with less than𝑘recorded interactions. The 𝑘-core strategy can proceed iteratively (Iterative𝑘-core) on both users and items until the 𝑘-core filtering condition is met, i.e., all the users and items have at least𝑘recorded interaction. Since reaching such condition might be intractable, Elliot allows specifying the maximum number of iterations (Iter-n-rounds). Finally, the Cold-Users filtering feature allows retaining cold-users only.

Elliot provides several prefiltering strategies. To enable Prefiltering operations, we can insert the corresponding block into our config file:

experiment:
  prefiltering:
    strategy: global_threshold|user_average|user_k_core|item_k_core|iterative_k_core|n_rounds_k_core|cold_users
    threshold: 3|average
    core: 5
    rounds: 2

In detail, Elliot provides eight main prefiltering approaches: global_threshold, user_average, user_k_core, item_k_core, iterative_k_core, n_rounds_k_core, cold_users.

global_threshold assumes a single system-wise threshold to filter out irrelevant transactions. global_threshold takes one mandatory parameter, threshold. threshold takes, as values, a float (ratings >= threshold will be kept), or the string average. With average, the system computes the global mean of the rating values and filters out all the ratings below.

experiment:
  prefiltering:
    strategy: global_threshold
    threshold: 3

experiment:
  prefiltering:
    strategy: global_threshold
    threshold: average

user_average has no parameters, and the system filters out the ratings below each user rating values mean.

experiment:
  prefiltering:
    strategy: user_average

user_k_core filters out all the users with a number of transactions lower than the given k core. It takes a parameter, core, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: user_k_core
    core: 5

item_k_core filters out all the items with a number of transactions lower than the given k core. It takes a parameter, core, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: item_k_core
    core: 5

iterative_k_core runs iteratively user_k_core, and item_k_core until the dataset is no further modified. It takes a parameter, core, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: iterative_k_core
    core: 5

n_rounds_k_core runs iteratively user_k_core, and item_k_core for a specified number of rounds. It takes the first parameter, core, where the user passes an int corresponding to the desired value. It takes the second parameter, rounds, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: n_rounds_k_core
    core: 5
    rounds: 2

cold_users filters out all the users with a number of interactions higher than a given threshold. It takes a parameter, threshold, where the user passes an int corresponding to the desired value.

experiment:
  prefiltering:
    strategy: cold_users
    threshold: 3

Splitting Data¶

If needed, the data is served to the Splitting module. In detail, Elliot provides (i)Temporal, (ii)Random, and (iii)Fix strategies. The Temporal strategy splits the user-item interactions based on the transaction timestamp, i.e., fixing the timestamp, find-ing the optimal one, or adopting a hold-out (HO) mechanism. The Random strategy includes hold-out (HO),𝐾-repeated hold-out(K-HO), and cross-validation (CV). Table 1 provides further configuration details. Finally, the Fix strategy exploits a precomputed splitting.

Elliot provides several splitting strategies. To enable the splitting operations, we can insert the corresponding section:

experiment:
  splitting:
    save_on_disk: True
    save_folder: this/is/the/path/
    test_splitting:
        strategy: fixed_timestamp|temporal_hold_out|random_subsampling|random_cross_validation
        timestamp: best|1609786061
        test_ratio: 0.2
        leave_n_out: 1
        folds: 5
    validation_splitting:
        strategy: fixed_timestamp|temporal_hold_out|random_subsampling|random_cross_validation
        timestamp: best|1609786061
        test_ratio: 0.2
        leave_n_out: 1
        folds: 5

Before deepening the splitting configurations, we can configure Elliot to save on disk the split files, once the splitting operation is completed.

To this extent, we can insert two fields into the section: save_on_disk, and save_folder.

save_on_disk enables the writing process, and save_folder specifies the system location where to save the split files:

experiment:
  splitting:
    save_on_disk: True
    save_folder: this/is/the/path/

Now, we can insert one (or two) specific subsections to detail the train/test, and the train/validation splitting via the corresponding fields: test_splitting, and validation_splitting. test_splitting is clearly mandatory, while validation_splitting is optional. Since the two subsections follow the same guidelines, here we detail test_splitting without loss of generality.

Elliot enables four splitting families: fixed_timestamp, temporal_hold_out, random_subsampling, random_cross_validation.

fixed_timestamp assumes that there will be a specific timestamp to split prior interactions (train) and future interactions. It takes the parameter timestamp, that can assume one of two possible kind of values: a long corresponding to a specific timestamp, or the string best computed following Anelli et al.

experiment:
  splitting:
    test_splitting:
        strategy: fixed_timestamp
        timestamp: 1609786061

experiment:
  splitting:
    test_splitting:
        strategy: fixed_timestamp
        timestamp: best

temporal_hold_out relies on a temporal split of user transactions. The split can be realized following two different approaches: a ratio-based and a leave-n-out-based approach. If we enable the test_ratio field with a float value, Elliot splits data retaining the last (100 * test_ratio) % of the user transactions for the test set. If we enable the leave_n_out field with an int value, Elliot retains the last leave_n_out transactions for the test set.

experiment:
  splitting:
    test_splitting:
        strategy: temporal_hold_out
        test_ratio: 0.2

experiment:
  splitting:
    test_splitting:
        strategy: temporal_hold_out
        leave_n_out: 1

random_subsampling generalizes random hold-out strategy. It takes a test_ratio parameter with a float value to define the train/test ratio for user-based hold-out splitting. Alternatively, it can take leave_n_out with an int value to define the number of transaction retained for the test set. Moreover, the splitting operation can be repeated enabling the folds field and passing an int. In that case, the overall splitting strategy corresponds to a user-based random subsampling strategy.

experiment:
  splitting:
    test_splitting:
        strategy: random_subsampling
        test_ratio: 0.2

experiment:
  splitting:
    test_splitting:
        strategy: random_subsampling
        test_ratio: 0.2
        folds: 5

experiment:
  splitting:
    test_splitting:
        strategy: random_subsampling
        leave_n_out: 1
        folds: 5

random_cross_validation adopts a k-folds cross-validation splitting strategy. It takes the parameter folds with an int value, that defines the overall number of folds to consider.

experiment:
  splitting:
    test_splitting:
        strategy: random_cross_validation
        folds: 5

Summary of the Recommendation Algorithms¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

All the recommendation models inherit from a common abstract class:

base_recommender_model.BaseRecommenderModel(…)

The majority of the recommendation models uses a Mixin:

recommender_utils_mixin.RecMixin()

Adversarial Learning

`adversarial.AMF.AMF.AMF`(data, config, …)	Adversarial Matrix Factorization
`adversarial.AMR.AMR.AMR`(data, config, …)	Adversarial Multimedia Recommender

Algebric

algebric.slope_one.slope_one.SlopeOne(data, …)

Slope One Predictors for Online Rating-Based Collaborative Filtering

Autoencoders

`autoencoders.dae.multi_dae.MultiDAE`(data, …)	Variational Autoencoders for Collaborative Filtering
`autoencoders.vae.multi_vae.MultiVAE`(data, …)	Variational Autoencoders for Collaborative Filtering

Content-Based

content_based.VSM.vector_space_model.VSM(…)

Vector Space Model

Generative Adversarial Networks (GANs)

`gan.IRGAN.irgan.IRGAN`(data, config, params, …)	IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models
`gan.CFGAN.cfgan.CFGAN`(data, config, params, …)	CFGAN: A Generic Collaborative Filtering Framework based on Generative Adversarial Networks

Graph-based

`graph_based.lightgcn.LightGCN.LightGCN`(data, …)	LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation
`graph_based.ngcf.NGCF.NGCF`(data, config, …)	Neural Graph Collaborative Filtering

Knowledge-aware

`knowledge_aware.kaHFM.ka_hfm.KaHFM`(data, …)	Knowledge-aware Hybrid Factorization Machines
`knowledge_aware.kaHFM_batch.kahfm_batch.KaHFMBatch`(…)	Knowledge-aware Hybrid Factorization Machines (Tensorflow Batch Variant)
`knowledge_aware.kahfm_embeddings.kahfm_embeddings.KaHFMEmbeddings`(…)	Knowledge-aware Hybrid Factorization Machines (Tensorflow Embedding-based Variant)

Latent Factor Models

`latent_factor_models.BPRMF.BPRMF.BPRMF`(data, …)	Bayesian Personalized Ranking with Matrix Factorization
`latent_factor_models.BPRMF_batch.BPRMF_batch.BPRMF_batch`(…)	Batch Bayesian Personalized Ranking with Matrix Factorization
`latent_factor_models.BPRSlim.bprslim.BPRSlim`(…)	BPR Sparse Linear Methods
`latent_factor_models.CML.CML.CML`(data, …)	Collaborative Metric Learning
`latent_factor_models.FFM.field_aware_factorization_machine.FFM`(…)	Field-aware Factorization Machines
`latent_factor_models.FISM.FISM.FISM`(data, …)	FISM: Factored Item Similarity Models
`latent_factor_models.FM.factorization_machine.FM`(…)	Factorization Machines
`latent_factor_models.FunkSVD.funk_svd.FunkSVD`(…)	For further details, please refer to the paper
`latent_factor_models.LogisticMF.logistic_matrix_factorization.LogisticMatrixFactorization`(…)	Logistic Matrix Factorization
`latent_factor_models.MF.matrix_factorization.MF`(…)	Matrix Factorization
`latent_factor_models.NonNegMF.non_negative_matrix_factorization.NonNegMF`(…)	Non-Negative Matrix Factorization
`latent_factor_models.PMF.probabilistic_matrix_factorization.PMF`(…)	Probabilistic Matrix Factorization
`latent_factor_models.PureSVD.pure_svd.PureSVD`(…)	For further details, please refer to the paper
`latent_factor_models.Slim.slim.Slim`(data, …)	Sparse Linear Methods
`latent_factor_models.SVDpp.svdpp.SVDpp`(data, …)	SVD++
`latent_factor_models.WRMF.wrmf.WRMF`(data, …)	Weighted XXX Matrix Factorization

Artificial Neural Networks

`neural.ConvMF.convolutional_matrix_factorization.ConvMF`(…)	Convolutional Matrix Factorization for Document Context-Aware Recommendation
`neural.ConvNeuMF.convolutional_neural_matrix_factorization.ConvNeuMF`(…)	Outer Product-based Neural Collaborative Filtering
`neural.DeepFM.deep_fm.DeepFM`(data, config, …)	DeepFM: A Factorization-Machine based Neural Network for CTR Prediction
`neural.DMF.deep_matrix_factorization.DMF`(…)	Deep Matrix Factorization Models for Recommender Systems.
`neural.GeneralizedMF.generalized_matrix_factorization.GMF`(…)	Neural Collaborative Filtering
`neural.ItemAutoRec.itemautorec.ItemAutoRec`(…)	AutoRec: Autoencoders Meet Collaborative Filtering (Item-based)
`neural.NAIS.nais.NAIS`(data, config, params, …)	NAIS: Neural Attentive Item Similarity Model for Recommendation
`neural.NeuMF.neural_matrix_factorization.NeuMF`(…)	Neural Collaborative Filtering
`neural.NFM.neural_fm.NFM`(data, config, …)	Neural Factorization Machines for Sparse Predictive Analytics
`neural.NPR.neural_personalized_ranking.NPR`(…)	Neural Personalized Ranking for Image Recommendation (Model without visual features)
`neural.UserAutoRec.userautorec.UserAutoRec`(…)	AutoRec: Autoencoders Meet Collaborative Filtering (User-based)
`neural.WideAndDeep.wide_and_deep.WideAndDeep`(…)	Wide & Deep Learning for Recommender Systems

Neighborhood-based Models

`NN.item_knn.item_knn.ItemKNN`(data, config, …)	Amazon.com recommendations: item-to-item collaborative filtering
`NN.user_knn.user_knn.UserKNN`(data, config, …)	GroupLens: An Open Architecture for Collaborative Filtering of Netnews
`NN.attribute_item_knn.attribute_item_knn.AttributeItemKNN`(…)	Attribute Item-kNN proposed in MyMediaLite Recommender System Library
`NN.attribute_user_knn.attribute_user_knn.AttributeUserKNN`(…)	Attribute User-kNN proposed in MyMediaLite Recommender System Library

Unpersonalized Recommenders

`unpersonalized.most_popular.most_popular.MostPop`(…)
`unpersonalized.random_recommender.Random.Random`(…)

Visual Models

`visual_recommenders.ACF.ACF.ACF`(data, …)	Attentive Collaborative Filtering: Multimedia Recommendation with Item- and Component-Level Attention
`visual_recommenders.DeepStyle.DeepStyle.DeepStyle`(…)	DeepStyle: Learning User Preferences for Visual Recommendation
`visual_recommenders.DVBPR.DVBPR.DVBPR`(data, …)	Visually-Aware Fashion Recommendation and Design with Generative Image Models
`visual_recommenders.VBPR.VBPR.VBPR`(data, …)	VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback
`visual_recommenders.VNPR.visual_neural_personalized_ranking.VNPR`(…)	Visual Neural Personalized Ranking for Image Recommendation

Hyperparameter Optimization¶

Elliot provides hyperparameter tuning optimization integrating the functionalities of the HyperOpt library and extending it with exhaustive grid search.

Before continuing, let us recall how to include a recommendation system into an experiment:

experiment:
  models:
    PMF:
      meta:
        hyper_max_evals: 20
        hyper_opt_alg: tpe
        validation_rate: 1
        verbose: True
        save_weights: True
        save_recs: True
        validation_metric: nDCG@10
      lr: 0.0025
      epochs: 2
      factors: 50
      batch_size: 512
      reg: 0.0025
      reg_b: 0
      gaussian_variance: 0.1

As we can observe, the meta section contains two fields that are related to hyperparameter optimization: hyper_max_evals, and hyper_opt_alg.

hyper_opt_alg is a string field that defines the hyperparameter tuning strategy

hyper_opt_alg can assume one of the following values: grid, tpe, *atpe, rand, mix, and anneal.

grid corresponds to exhaustive grid search

tpe stands for Tree of Parzen Estimators, a type of Bayesian Optimization, see the paper

atpe stands for Adaptive Tree of Parzen Estimators

rand stands for random sampling in the search space

mix stands for mixture of search algorithms

anneal stands for simulated annealing

hyper_max_evals is an int field that, where applicable (all strategies but grid), defines the number of samples to consider for hyperparameter evaluation

Once we choose the search strategy, we need to define the search space. To this end, Elliot provides two alternatives: a value list, and a function-parameters pair.

In the former case, we just need to provide a list of values to the parameter we want to optimize:

experiment:
  models:
    PMF:
      meta:
        hyper_max_evals: 20
        hyper_opt_alg: tpe
      lr:                   0.0025
      epochs:               2
      factors:              50
      batch_size:           512
      reg:                  [0.0025, 0.005, 0.01]
      reg_b:                0
      gaussian_variance:    0.1

In the latter case, we can choose among the search space functions provided by HyperOpt: choice, randint, uniform, quniform, loguniform, qloguniform, normal, qnormal, lognormal, qlognormal. Each function and its parameters are documented at the page in the section Parameters Expression.

Note that the label argument is internal and DO NOT have to provide it.

To teach Elliot to sample from any of these search spaces is straightforward: we pass to the parameter a list in which the first element is the function name, and the others are the parameter values.

An example of the syntax to define a search with loguniform for the learning rate parameter (lr) is:

experiment:
  models:
    PMF:
      meta:
        hyper_max_evals: 20
        hyper_opt_alg: tpe
      lr:                   [loguniform, -10, -1]
      epochs:               2
      factors:              50
      batch_size:           512
      reg:                  [0.0025, 0.005, 0.01]
      reg_b:                0
      gaussian_variance:    0.1

Finally, Elliot provides a shortcut to perform an exhaustive grid search. We can avoid inserting hyper_opt_alg and hyper_max_evals fields and we directly insert the lists of possible values for the parameters to optimize:

experiment:
  models:
    PMF:
      meta:
        validation_rate: 1
        verbose: True
        save_weights: True
        save_recs: True
        validation_metric: nDCG@10
      lr:                   [0.0025, 0.005, 0.01]
      epochs:               50
      factors:              [10, 50, 100]
      batch_size:           512
      reg:                  [0.0025, 0.005, 0.01]
      reg_b:                0
      gaussian_variance:    0.1

In this case, Elliot recognizes that hyperparameter optimization is needed and automatically performs the grid search.

Create a new Recommendation Model¶

Work in progress

Recommendation Models¶

Adversarial Learning¶

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed id porta mi. Proin luctus sapien ut mauris facilisis, in faucibus quam cursus. Pellentesque eget lacus eros. Aenean eget molestie magna. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos himenaeos. Nam dapibus erat at scelerisque facilisis. Cras diam dolor, viverra et ipsum ac, ultrices lacinia eros.

Summary¶

`AMF.AMF.AMF`(data, config, params, *args, …)	Adversarial Matrix Factorization
`AMR.AMR.AMR`(data, config, params, *args, …)	Adversarial Multimedia Recommender

AMF¶

class elliot.recommender.adversarial.AMF.AMF.AMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Adversarial Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factor
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
eps – Perturbation Budget
l_adv – Adversarial regularization coefficient
adversarial_epochs – Adversarial epochs

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 200
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    eps: 0.1
    l_adv: 0.001
    adversarial_epochs: 10

AMR¶

class elliot.recommender.adversarial.AMR.AMR.AMR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Adversarial Multimedia Recommender

For further details, please refer to the paper

Parameters

factors – Number of latent factor
factors_d – Image-feature dimensionality
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_e – Regularization coefficient of image matrix embedding
eps – Perturbation Budget
l_adv – Adversarial regularization coefficient
adversarial_epochs – Adversarial epochs

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AMR:
    meta:
      save_recs: True
    epochs: 10
    factors: 200
    factors_d: 20
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    l_e: 0.1
    eps: 0.1
    l_adv: 0.001
    adversarial_epochs: 5

Algebric¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

slope_one.slope_one.SlopeOne(data, config, …)

Slope One Predictors for Online Rating-Based Collaborative Filtering

SlopeOne¶

class elliot.recommender.algebric.slope_one.slope_one.SlopeOne(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Slope One Predictors for Online Rating-Based Collaborative Filtering

For further details, please refer to the paper

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  SlopeOne:
    meta:
      save_recs: True

Autoencoders¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

`dae.multi_dae.MultiDAE`(data, config, params, …)	Variational Autoencoders for Collaborative Filtering
`vae.multi_vae.MultiVAE`(data, config, params, …)	Variational Autoencoders for Collaborative Filtering

MultiDAE¶

class elliot.recommender.autoencoders.dae.multi_dae.MultiDAE(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Variational Autoencoders for Collaborative Filtering

For further details, please refer to the paper

Parameters

intermediate_dim – Number of intermediate dimension
latent_dim – Number of latent factors
reg_lambda – Regularization coefficient
lr – Learning rate
dropout_pkeep – Dropout probaility

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  MultiDAE:
    meta:
      save_recs: True
    epochs: 10
    intermediate_dim: 600
    latent_dim: 200
    reg_lambda: 0.01
    lr: 0.001
    dropout_pkeep: 1

MultiVAE¶

class elliot.recommender.autoencoders.vae.multi_vae.MultiVAE(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Variational Autoencoders for Collaborative Filtering

For further details, please refer to the paper

Parameters

intermediate_dim – Number of intermediate dimension
latent_dim – Number of latent factors
reg_lambda – Regularization coefficient
lr – Learning rate
dropout_pkeep – Dropout probaility

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  MultiVAE:
    meta:
      save_recs: True
    epochs: 10
    intermediate_dim: 600
    latent_dim: 200
    reg_lambda: 0.01
    lr: 0.001
    dropout_pkeep: 1

Content-Based¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

VSM.vector_space_model.VSM(data, config, …)

Vector Space Model

VSM¶

class elliot.recommender.content_based.VSM.vector_space_model.VSM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Vector Space Model

For further details, please refer to the paper and the paper

Parameters

similarity – Similarity metric
user_profile –
item_profile –

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  VSM:
    meta:
      save_recs: True
    similarity: cosine
    user_profile: binary
    item_profile: binary

Generative Adversarial Networks (GANs)¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

`IRGAN.irgan.IRGAN`(data, config, params, …)	IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models
`CFGAN.cfgan.CFGAN`(data, config, params, …)	CFGAN: A Generic Collaborative Filtering Framework based on Generative Adversarial Networks

IRGAN¶

class elliot.recommender.gan.IRGAN.irgan.IRGAN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models

For further details, please refer to the paper

Parameters

factors – Number of latent factor
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_gan – Adversarial regularization coefficient
predict_model – Specification of the model to generate the recommendation (Generator/ Discriminator)
g_epochs – Number of epochs to train the generator for each IRGAN step
d_epochs – Number of epochs to train the discriminator for each IRGAN step
g_pretrain_epochs – Number of epochs to pre-train the generator
d_pretrain_epochs – Number of epochs to pre-train the discriminator
sample_lambda – Temperature Parameters

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  IRGAN:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    l_gan: 0.001
    predict_model: generator
    g_epochs: 5
    d_epochs: 1
    g_pretrain_epochs: 10
    d_pretrain_epochs: 10
    sample_lambda: 0.2

CFGAN¶

class elliot.recommender.gan.CFGAN.cfgan.CFGAN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

CFGAN: A Generic Collaborative Filtering Framework based on Generative Adversarial Networks

For further details, please refer to the paper

Parameters

factors – Number of latent factor
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_gan – Adversarial regularization coefficient
g_epochs – Number of epochs to train the generator for each IRGAN step
d_epochs – Number of epochs to train the discriminator for each IRGAN step
s_zr – Sampling parameter of zero-reconstruction
s_pm – Sampling parameter of partial-masking

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  CFGAN:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    l_gan: 0.001
    g_epochs: 5
    d_epochs: 1
    s_zr: 0.001
    s_pm: 0.001

Graph-based¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

`lightgcn.LightGCN.LightGCN`(data, config, …)	LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation
`ngcf.NGCF.NGCF`(data, config, params, *args, …)	Neural Graph Collaborative Filtering

LightGCN¶

class elliot.recommender.graph_based.lightgcn.LightGCN.LightGCN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
l_w – Regularization coefficient
n_layers – Number of embedding propagation layers
n_fold – Number of folds to split the adjacency matrix into sub-matrices and ease the computation

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  LightGCN:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 64
    batch_size: 256
    l_w: 0.1
    n_layers: 1
    n_fold: 5

NGCF¶

class elliot.recommender.graph_based.ngcf.NGCF.NGCF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Graph Collaborative Filtering

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
l_w – Regularization coefficient
weight_size – Tuple with number of units for each embedding propagation layer
node_dropout – Tuple with dropout rate for each node
message_dropout – Tuple with dropout rate for each embedding propagation layer
n_fold – Number of folds to split the adjacency matrix into sub-matrices and ease the computation

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NGCF:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 64
    batch_size: 256
    l_w: 0.1
    weight_size: (64,)
    node_dropout: ()
    message_dropout: (0.1,)
    n_fold: 5

Knowledge-aware¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

`kaHFM.ka_hfm.KaHFM`(data, config, params, …)	Knowledge-aware Hybrid Factorization Machines
`kaHFM_batch.kahfm_batch.KaHFMBatch`(data, …)	Knowledge-aware Hybrid Factorization Machines (Tensorflow Batch Variant)
`kahfm_embeddings.kahfm_embeddings.KaHFMEmbeddings`(…)	Knowledge-aware Hybrid Factorization Machines (Tensorflow Embedding-based Variant)

KaHFM¶

class elliot.recommender.knowledge_aware.kaHFM.ka_hfm.KaHFM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Knowledge-aware Hybrid Factorization Machines

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “How to Make Latent Factors Interpretable by Feeding Factorization Machines with Knowledge Graphs”, ISWC 2019 Best student Research Paper For further details, please refer to the paper

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “Semantic Interpretation of Top-N Recommendations”, IEEE TKDE 2020 For further details, please refer to the paper

Parameters

lr – learning rate (default: 0.05)
bias_regularization – Bias regularization (default: 0)
user_regularization – User regularization (default: 0.0025)
positive_item_regularization – regularization for positive (experienced) items (default: 0.0025)
negative_item_regularization – regularization for unknown items (default: 0.00025)
update_negative_item_factors – Boolean to update negative item factors (default: True)
update_users – Boolean to update user factors (default: True)
update_items – Boolean to update item factors (default: True)
update_bias – Boolean to update bias value (default: True)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  KaHFM:
    meta:
      hyper_max_evals: 20
      hyper_opt_alg: tpe
      validation_rate: 1
      verbose: True
      save_weights: True
      save_recs: True
      validation_metric: nDCG@10
    epochs: 100
    batch_size: -1
    lr: 0.05
    bias_regularization: 0
    user_regularization: 0.0025
    positive_item_regularization: 0.0025
    negative_item_regularization: 0.00025
    update_negative_item_factors: True
    update_users: True
    update_items: True
    update_bias: True

KaHFM Batch¶

class elliot.recommender.knowledge_aware.kaHFM_batch.kahfm_batch.KaHFMBatch(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Knowledge-aware Hybrid Factorization Machines (Tensorflow Batch Variant)

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “How to Make Latent Factors Interpretable by Feeding Factorization Machines with Knowledge Graphs”, ISWC 2019 Best student Research Paper For further details, please refer to the paper

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “Semantic Interpretation of Top-N Recommendations”, IEEE TKDE 2020 For further details, please refer to the paper

Parameters

lr – learning rate (default: 0.0001)
l_w – Weight regularization (default: 0.005)
l_b – Bias regularization (default: 0)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  KaHFMBatch:
    meta:
      hyper_max_evals: 20
      hyper_opt_alg: tpe
      validation_rate: 1
      verbose: True
      save_weights: True
      save_recs: True
      validation_metric: nDCG@10
    epochs: 100
    batch_size: -1
    lr: 0.0001
    l_w: 0.005
    l_b: 0

KaHFM Embeddings¶

class elliot.recommender.knowledge_aware.kahfm_embeddings.kahfm_embeddings.KaHFMEmbeddings(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Knowledge-aware Hybrid Factorization Machines (Tensorflow Embedding-based Variant)

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “How to Make Latent Factors Interpretable by Feeding Factorization Machines with Knowledge Graphs”, ISWC 2019 Best student Research Paper For further details, please refer to the paper

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “Semantic Interpretation of Top-N Recommendations”, IEEE TKDE 2020 For further details, please refer to the paper

Parameters

lr – learning rate (default: 0.0001)
l_w – Weight regularization (default: 0.005)
l_b – Bias regularization (default: 0)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  KaHFMEmbeddings:
    meta:
      hyper_max_evals: 20
      hyper_opt_alg: tpe
      validation_rate: 1
      verbose: True
      save_weights: True
      save_recs: True
      validation_metric: nDCG@10
    epochs: 100
    batch_size: -1
    lr: 0.0001
    l_w: 0.005
    l_b: 0

Latent Factor Models¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

`BPRMF.BPRMF.BPRMF`(data, config, params, …)	Bayesian Personalized Ranking with Matrix Factorization
`BPRMF_batch.BPRMF_batch.BPRMF_batch`(data, …)	Batch Bayesian Personalized Ranking with Matrix Factorization
`BPRSlim.bprslim.BPRSlim`(data, config, …)	BPR Sparse Linear Methods
`CML.CML.CML`(data, config, params, *args, …)	Collaborative Metric Learning
`FFM.field_aware_factorization_machine.FFM`(…)	Field-aware Factorization Machines
`FISM.FISM.FISM`(data, config, params, *args, …)	FISM: Factored Item Similarity Models
`FM.factorization_machine.FM`(data, config, …)	Factorization Machines
`FunkSVD.funk_svd.FunkSVD`(data, config, …)	For further details, please refer to the paper
`LogisticMF.logistic_matrix_factorization.LogisticMatrixFactorization`(…)	Logistic Matrix Factorization
`MF.matrix_factorization.MF`(data, config, …)	Matrix Factorization
`NonNegMF.non_negative_matrix_factorization.NonNegMF`(…)	Non-Negative Matrix Factorization
`PMF.probabilistic_matrix_factorization.PMF`(…)	Probabilistic Matrix Factorization
`PureSVD.pure_svd.PureSVD`(data, config, …)	For further details, please refer to the paper
`Slim.slim.Slim`(data, config, params, *args, …)	Sparse Linear Methods
`SVDpp.svdpp.SVDpp`(data, config, params, …)	SVD++
`WRMF.wrmf.WRMF`(data, config, params, *args, …)	Weighted XXX Matrix Factorization

BPRMF¶

class elliot.recommender.latent_factor_models.BPRMF.BPRMF.BPRMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Bayesian Personalized Ranking with Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
bias_regularization – Regularization coefficient for the bias
user_regularization – Regularization coefficient for user latent factors
positive_item_regularization – Regularization coefficient for positive item latent factors
negative_item_regularization – Regularization coefficient for negative item latent factors
update_negative_item_factors –
update_users –
update_items –
update_bias –

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  BPRMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    bias_regularization: 0
    user_regularization: 0.0025
    positive_item_regularization: 0.0025
    negative_item_regularization: 0.0025
    update_negative_item_factors: True
    update_users: True
    update_items: True
    update_bias: True

BPRMF_batch¶

class elliot.recommender.latent_factor_models.BPRMF_batch.BPRMF_batch.BPRMF_batch(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Batch Bayesian Personalized Ranking with Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
l_w – Regularization coefficient for latent factors
l_b – Regularization coefficient for bias

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  BPRMF_batch:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.1
    l_b: 0.001

BPRSlim¶

class elliot.recommender.latent_factor_models.BPRSlim.bprslim.BPRSlim(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

BPR Sparse Linear Methods

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
lj_reg – Regularization coefficient for positive items
li_reg – Regularization coefficient for negative items

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    lj_reg: 0.001
    li_reg: 0.1

CML¶

class elliot.recommender.latent_factor_models.CML.CML.CML(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Collaborative Metric Learning

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
l_w – Regularization coefficient for latent factors
l_b – Regularization coefficient for bias
margin – Safety margin size

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  CML:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.001
    l_b: 0.001
    margin: 0.5

FFM¶

class elliot.recommender.latent_factor_models.FFM.field_aware_factorization_machine.FFM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Field-aware Factorization Machines

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  FFM:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1

FISM¶

class elliot.recommender.latent_factor_models.FISM.FISM.FISM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

FISM: Factored Item Similarity Models

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
l_w – Regularization coefficient for latent factors
l_b – Regularization coefficient for bias
alpha – Alpha parameter (a value between 0 and 1)
neg_ratio –

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  FISM:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.001
    l_b: 0.001
    alpha: 0.5
    neg_ratio:

FM¶

class elliot.recommender.latent_factor_models.FM.factorization_machine.FM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Factorization Machines

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  FM:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1

FunkSVD¶

class elliot.recommender.latent_factor_models.FunkSVD.funk_svd.FunkSVD(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
reg_w – Regularization coefficient for latent factors
reg_b – Regularization coefficient for bias

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  FunkSVD:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg_w: 0.1
    reg_b: 0.001

LogisticMatrixFactorization¶

class elliot.recommender.latent_factor_models.LogisticMF.logistic_matrix_factorization.LogisticMatrixFactorization(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Logistic Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
reg – Regularization coefficient
alpha – Parameter for confidence estimation

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  LogisticMatrixFactorization:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1
    alpha: 0.5

NonNegMF¶

class elliot.recommender.latent_factor_models.NonNegMF.non_negative_matrix_factorization.NonNegMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Non-Negative Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NonNegMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1

PMF¶

class elliot.recommender.latent_factor_models.PMF.probabilistic_matrix_factorization.PMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Probabilistic Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
reg – Regularization coefficient
gaussian_variance – Variance of the Gaussian distribution

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  PMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 50
    lr: 0.001
    reg: 0.0025
    gaussian_variance: 0.1

PureSVD¶

class elliot.recommender.latent_factor_models.PureSVD.pure_svd.PureSVD(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

For further details, please refer to the paper

Parameters

factors – Number of latent factors
seed – Random seed

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  PureSVD:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    seed: 42

Slim¶

class elliot.recommender.latent_factor_models.Slim.slim.Slim(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Sparse Linear Methods

For further details, please refer to the paper

Parameters

l1_ratio –
alpha –

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  Slim:
    meta:
      save_recs: True
    epochs: 10
    l1_ratio: 0.001
    alpha: 0.001

SVDpp¶

class elliot.recommender.latent_factor_models.SVDpp.svdpp.SVDpp(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

SVD++

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
reg_w – Regularization coefficient for latent factors
reg_b – Regularization coefficient for bias

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  SVDpp:
    meta:
      save_recs: True
    epochs: 10
    factors: 50
    lr: 0.001
    reg_w: 0.1
    reg_b: 0.001

WRMF¶

class elliot.recommender.latent_factor_models.WRMF.wrmf.WRMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Weighted XXX Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
alpha –
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  WRMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 50
    alpha: 1
    reg: 0.1

Artificial Neural Networks¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

`ConvMF.convolutional_matrix_factorization.ConvMF`(…)	Convolutional Matrix Factorization for Document Context-Aware Recommendation
`ConvNeuMF.convolutional_neural_matrix_factorization.ConvNeuMF`(…)	Outer Product-based Neural Collaborative Filtering
`DeepFM.deep_fm.DeepFM`(data, config, params, …)	DeepFM: A Factorization-Machine based Neural Network for CTR Prediction
`DMF.deep_matrix_factorization.DMF`(data, …)	Deep Matrix Factorization Models for Recommender Systems.
`GeneralizedMF.generalized_matrix_factorization.GMF`(…)	Neural Collaborative Filtering
`ItemAutoRec.itemautorec.ItemAutoRec`(data, …)	AutoRec: Autoencoders Meet Collaborative Filtering (Item-based)
`NAIS.nais.NAIS`(data, config, params, *args, …)	NAIS: Neural Attentive Item Similarity Model for Recommendation
`NeuMF.neural_matrix_factorization.NeuMF`(…)	Neural Collaborative Filtering
`NFM.neural_fm.NFM`(data, config, params, …)	Neural Factorization Machines for Sparse Predictive Analytics
`NPR.neural_personalized_ranking.NPR`(data, …)	Neural Personalized Ranking for Image Recommendation (Model without visual features)
`UserAutoRec.userautorec.UserAutoRec`(data, …)	AutoRec: Autoencoders Meet Collaborative Filtering (User-based)
`WideAndDeep.wide_and_deep.WideAndDeep`(data, …)	Wide & Deep Learning for Recommender Systems

ConvMF¶

class elliot.recommender.neural.ConvMF.convolutional_matrix_factorization.ConvMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Convolutional Matrix Factorization for Document Context-Aware Recommendation

For further details, please refer to the paper

Parameters

embedding_size – Embedding dimension
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
cnn_channels – List of channels
cnn_kernels – List of kernels
cnn_strides – List of strides
dropout_prob – Dropout probability applied on the convolutional layers

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ConvMF:
    meta:
      save_recs: True
    epochs: 10
    embedding_size: 100
    lr: 0.001
    l_w: 0.005
    l_b: 0.0005
    cnn_channels: (1, 32, 32)
    cnn_kernels: (2,2)
    cnn_strides: (2,2)
    dropout_prob: 0

ConvNeuMF¶

class elliot.recommender.neural.ConvNeuMF.convolutional_neural_matrix_factorization.ConvNeuMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Outer Product-based Neural Collaborative Filtering

For further details, please refer to the paper

Parameters

embedding_size – Embedding dimension
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
cnn_channels – List of channels
cnn_kernels – List of kernels
cnn_strides – List of strides
dropout_prob – Dropout probability applied on the convolutional layers

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ConvNeuMF:
    meta:
      save_recs: True
    epochs: 10
    embedding_size: 100
    lr: 0.001
    l_w: 0.005
    l_b: 0.0005
    cnn_channels: (1, 32, 32)
    cnn_kernels: (2,2)
    cnn_strides: (2,2)
    dropout_prob: 0

DeepFM¶

class elliot.recommender.neural.DeepFM.deep_fm.DeepFM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

DeepFM: A Factorization-Machine based Neural Network for CTR Prediction

For further details, please refer to the paper

Parameters

factors – Number of factors dimension
lr – Learning rate
l_w – Regularization coefficient
hidden_neurons – List of units for each layer
hidden_activations – List of activation functions

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  DeepFM:
    meta:
      save_recs: True
    epochs: 10
    factors: 100
    lr: 0.001
    l_w: 0.0001
    hidden_neurons: (64,32)
    hidden_activations: ('relu','relu')

DMF¶

class elliot.recommender.neural.DMF.deep_matrix_factorization.DMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Deep Matrix Factorization Models for Recommender Systems.

For further details, please refer to the paper

Parameters

lr – Learning rate
reg – Regularization coefficient
user_mlp – List of units for each layer
item_mlp – List of activation functions
similarity – Number of factors dimension

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  DMF:
    meta:
      save_recs: True
    epochs: 10
    lr: 0.0001
    reg: 0.001
    user_mlp: (64,32)
    item_mlp: (64,32)
    similarity: cosine

GMF¶

class elliot.recommender.neural.GeneralizedMF.generalized_matrix_factorization.GMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Collaborative Filtering

For further details, please refer to the paper

Parameters

mf_factors – Number of latent factors
lr – Learning rate
is_edge_weight_train – Whether the training uses edge weighting

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  GMF:
    meta:
      save_recs: True
    epochs: 10
    mf_factors: 10
    lr: 0.001
    is_edge_weight_train: True

ItemAutoRec¶

class elliot.recommender.neural.ItemAutoRec.itemautorec.ItemAutoRec(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

AutoRec: Autoencoders Meet Collaborative Filtering (Item-based)

For further details, please refer to the paper

Parameters

hidden_neuron – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ItemAutoRec:
    meta:
      save_recs: True
    epochs: 10
    hidden_neuron: 500
    lr: 0.0001
    l_w: 0.001

NAIS¶

class elliot.recommender.neural.NAIS.nais.NAIS(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

NAIS: Neural Attentive Item Similarity Model for Recommendation

For further details, please refer to the paper

Parameters

factors – Number of latent factors
algorithm – Type of user-item factor operation (‘product’, ‘concat’)
weight_size – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient
l_b – Bias regularization coefficient
alpha – Attention factor
beta – Smoothing exponent
neg_ratio – Ratio of negative sampled items, e.g., 0 = no items, 1 = all un-rated items

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NAIS:
    meta:
      save_recs: True
    factors: 100
    algorithm: concat
    weight_size: 32
    lr: 0.001
    l_w: 0.001
    l_b: 0.001
    alpha: 0.5
    beta: 0.5
    neg_ratio: 0.5

NeuMF¶

class elliot.recommender.neural.NeuMF.neural_matrix_factorization.NeuMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Collaborative Filtering

For further details, please refer to the paper

Parameters

mf_factors – Number of MF latent factors
mlp_factors – Number of MLP latent factors
mlp_hidden_size – List of units for each layer
lr – Learning rate
dropout – Dropout rate
is_mf_train – Whether to train the MF embeddings
is_mlp_train – Whether to train the MLP layers

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NeuMF:
    meta:
      save_recs: True
    epochs: 10
    mf_factors: 10
    mlp_factors: 10
    mlp_hidden_size: (64,32)
    lr: 0.001
    dropout: 0.0
    is_mf_train: True
    is_mlp_train: True

NFM¶

class elliot.recommender.neural.NFM.neural_fm.NFM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Factorization Machines for Sparse Predictive Analytics

For further details, please refer to the paper

Parameters

factors – Number of factors dimension
lr – Learning rate
l_w – Regularization coefficient
hidden_neurons – List of units for each layer
hidden_activations – List of activation functions

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NFM:
    meta:
      save_recs: True
    epochs: 10
    factors: 100
    lr: 0.001
    l_w: 0.0001
    hidden_neurons: (64,32)
    hidden_activations: ('relu','relu')

NPR¶

class elliot.recommender.neural.NPR.neural_personalized_ranking.NPR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Personalized Ranking for Image Recommendation (Model without visual features)

For further details, please refer to the paper

Parameters

mf_factors – Number of MF latent factors
mlp_hidden_size – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient
dropout – Dropout rate

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NPR:
    meta:
      save_recs: True
    epochs: 10
    mf_factors: 100
    mlp_hidden_size:  (64,32)
    lr: 0.001
    l_w: 0.001
    dropout: 0.45

UserAutoRec¶

class elliot.recommender.neural.UserAutoRec.userautorec.UserAutoRec(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

AutoRec: Autoencoders Meet Collaborative Filtering (User-based)

For further details, please refer to the paper

Parameters

hidden_neuron – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  UserAutoRec:
    meta:
      save_recs: True
    epochs: 10
    hidden_neuron: 500
    lr: 0.0001
    l_w: 0.001

WideAndDeep¶

class elliot.recommender.neural.WideAndDeep.wide_and_deep.WideAndDeep(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Wide & Deep Learning for Recommender Systems

(For now, available with knowledge-aware features)

For further details, please refer to the paper

Parameters

factors – Number of latent factors
mlp_hidden_size – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient
l_b – Bias Regularization Coefficient
dropout_prob – Dropout rate

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NPR:
    meta:
      save_recs: True
    epochs: 10
    factors: 50
    mlp_hidden_size: (32, 32, 1)
    lr: 0.001
    l_w: 0.005
    l_b: 0.0005
    dropout_prob: 0.0

Neighborhood-based Models¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

`item_knn.item_knn.ItemKNN`(data, config, …)	Amazon.com recommendations: item-to-item collaborative filtering
`user_knn.user_knn.UserKNN`(data, config, …)	GroupLens: An Open Architecture for Collaborative Filtering of Netnews
`attribute_item_knn.attribute_item_knn.AttributeItemKNN`(…)	Attribute Item-kNN proposed in MyMediaLite Recommender System Library
`attribute_user_knn.attribute_user_knn.AttributeUserKNN`(…)	Attribute User-kNN proposed in MyMediaLite Recommender System Library

ItemKNN¶

class elliot.recommender.NN.item_knn.item_knn.ItemKNN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Amazon.com recommendations: item-to-item collaborative filtering

For further details, please refer to the paper

Parameters

neighbors – Number of item neighbors
similarity – Similarity function
implementation – Implementation type (‘aiolli’, ‘classical’)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ItemKNN:
    meta:
      save_recs: True
    neighbors: 40
    similarity: cosine
    implementation: aiolli

UserKNN¶

class elliot.recommender.NN.user_knn.user_knn.UserKNN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

GroupLens: An Open Architecture for Collaborative Filtering of Netnews

For further details, please refer to the paper

Parameters

neighbors – Number of item neighbors
similarity – Similarity function
implementation – Implementation type (‘aiolli’, ‘classical’)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  UserKNN:
    meta:
      save_recs: True
    neighbors: 40
    similarity: cosine
    implementation: aiolli

AttributeItemKNN¶

class elliot.recommender.NN.attribute_item_knn.attribute_item_knn.AttributeItemKNN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Attribute Item-kNN proposed in MyMediaLite Recommender System Library

For further details, please refer to the paper

Parameters

neighbors – Number of item neighbors
similarity – Similarity function

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AttributeItemKNN:
    meta:
      save_recs: True
    neighbors: 40
    similarity: cosine

AttributeUserKNN¶

class elliot.recommender.NN.attribute_user_knn.attribute_user_knn.AttributeUserKNN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Attribute User-kNN proposed in MyMediaLite Recommender System Library

For further details, please refer to the paper

Parameters

neighbors – Number of item neighbors
similarity – Similarity function
profile – Profile type (‘binary’, ‘tfidf’)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AttributeUserKNN:
    meta:
      save_recs: True
    neighbors: 40
    similarity: cosine
    profile: binary

Unpersonalized Recommenders¶

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed id porta mi. Proin luctus sapien ut mauris facilisis, in faucibus quam cursus. Pellentesque eget lacus eros. Aenean eget molestie magna. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos himenaeos. Nam dapibus erat at scelerisque facilisis. Cras diam dolor, viverra et ipsum ac, ultrices lacinia eros.

Summary¶

`most_popular.most_popular.MostPop`(data, …)
`random_recommender.Random.Random`(data, …)

Most Popular¶

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed id porta mi. Proin luctus sapien ut mauris facilisis, in faucibus quam cursus. Pellentesque eget lacus eros. Aenean eget molestie magna. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos himenaeos. Nam dapibus erat at scelerisque facilisis. Cras diam dolor, viverra et ipsum ac, ultrices lacinia eros.

class elliot.recommender.unpersonalized.most_popular.most_popular.MostPop(data, config, params, *args, **kwargs)[source]¶: Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

To include the recommendation model, add it to the config file adopting the following pattern:¶

models:
  MostPop:
    meta:
      save_recs: True

Random Recommender¶

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed id porta mi. Proin luctus sapien ut mauris facilisis, in faucibus quam cursus. Pellentesque eget lacus eros. Aenean eget molestie magna. Class aptent taciti sociosqu ad litora torquent per conubia nostra, per inceptos himenaeos. Nam dapibus erat at scelerisque facilisis. Cras diam dolor, viverra et ipsum ac, ultrices lacinia eros.

class elliot.recommender.unpersonalized.random_recommender.Random.Random(data, config, params, *args, **kwargs)[source]¶: Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

To include the recommendation model, add it to the config file adopting the following pattern:¶

models:
  Random:
    meta:
      save_recs: True
    random_seed: 42

Visual Models¶

Elliot integrates, to date, 50 recommendation models partitioned into two sets. The first set includes 38 popular models implemented in at least two of frameworks reviewed in this work (i.e., adopting a framework-wise popularity notion).

Summary¶

`ACF.ACF.ACF`(data, config, params, *args, …)	Attentive Collaborative Filtering: Multimedia Recommendation with Item- and Component-Level Attention
`DeepStyle.DeepStyle.DeepStyle`(data, config, …)	DeepStyle: Learning User Preferences for Visual Recommendation
`DVBPR.DVBPR.DVBPR`(data, config, params, …)	Visually-Aware Fashion Recommendation and Design with Generative Image Models
`VBPR.VBPR.VBPR`(data, config, params, *args, …)	VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback
`VNPR.visual_neural_personalized_ranking.VNPR`(…)	Visual Neural Personalized Ranking for Image Recommendation
`elliot.recommender.adversarial.AMR.AMR.AMR`(…)	Adversarial Multimedia Recommender

ACF¶

class elliot.recommender.visual_recommenders.ACF.ACF.ACF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Attentive Collaborative Filtering: Multimedia Recommendation with Item- and Component-Level Attention

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
l_w – Regularization coefficient
layers_component – Tuple with number of units for each attentive layer (component-level)
layers_item – Tuple with number of units for each attentive layer (item-level)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ACF:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 100
    batch_size: 128
    l_w: 0.000025
    layers_component: (64, 1)
    layers_item: (64, 1)

DeepStyle¶

class elliot.recommender.visual_recommenders.DeepStyle.DeepStyle.DeepStyle(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

DeepStyle: Learning User Preferences for Visual Recommendation

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
l_w – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  DeepStyle:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 100
    batch_size: 128
    l_w: 0.000025

DVBPR¶

class elliot.recommender.visual_recommenders.DVBPR.DVBPR.DVBPR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Visually-Aware Fashion Recommendation and Design with Generative Image Models

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
lambda_1 – Regularization coefficient
lambda_2 – CNN regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  DVBPR:
    meta:
      save_recs: True
    lr: 0.0001
    epochs: 50
    factors: 100
    batch_size: 128
    lambda_1: 0.0001
    lambda_2: 1.0

VBPR¶

class elliot.recommender.visual_recommenders.VBPR.VBPR.VBPR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
factors_d – Dimension of visual factors
batch_size – Batch size
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_e – Regularization coefficient of projection matrix

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  VBPR:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 100
    factors_d: 20
    batch_size: 128
    l_w: 0.000025
    l_b: 0
    l_e: 0.002

VNPR¶

class elliot.recommender.visual_recommenders.VNPR.visual_neural_personalized_ranking.VNPR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Visual Neural Personalized Ranking for Image Recommendation

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
mf_factors: – Number of latent factors for Matrix Factorization:
mlp_hidden_size – Tuple with number of units for each multi-layer perceptron layer
prob_keep_dropout – Dropout rate for multi-layer perceptron
batch_size – Batch size
l_w – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  VNPR:
    meta:
      save_recs: True
    lr: 0.001
    epochs: 50
    mf_factors: 10
    mlp_hidden_size: (32, 1)
    prob_keep_dropout: 0.2
    batch_size: 64
    l_w: 0.001

AMR¶

class elliot.recommender.adversarial.AMR.AMR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Adversarial Multimedia Recommender

For further details, please refer to the paper

Parameters

factors – Number of latent factor
factors_d – Image-feature dimensionality
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_e – Regularization coefficient of image matrix embedding
eps – Perturbation Budget
l_adv – Adversarial regularization coefficient
adversarial_epochs – Adversarial epochs

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AMR:
    meta:
      save_recs: True
    epochs: 10
    factors: 200
    factors_d: 20
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    l_e: 0.1
    eps: 0.1
    l_adv: 0.001
    adversarial_epochs: 5

Metrics¶

Elliot provides 36 evaluation metrics, partitioned into seven families: Accuracy, Rating-Error, Coverage, Novelty, Diversity, Bias, and Fairness. It is worth mentioning that Elliot is the framework that exposes both the largest number of metrics and the only one considering bias and fairness measures. Moreover, the user can choose any metric to drive the model selection and the tuning.

All the metrics inherit from a common abstract class:

base_metric.BaseMetric(recommendations, …)

This class represents the implementation of the Precision recommendation metric.

Accuracy

`accuracy.AUC.auc.AUC`(recommendations, …)	Area Under the Curve
`accuracy.AUC.gauc.GAUC`(recommendations, …)	Group Area Under the Curve
`accuracy.AUC.lauc.LAUC`(recommendations, …)	Limited Area Under the Curve
`accuracy.DSC.dsc.DSC`(recommendations, …)	Sørensen–Dice coefficient
`accuracy.f1.f1.F1`(recommendations, config, …)	F-Measure
`accuracy.f1.extended_f1.ExtendedF1`(…)	Extended F-Measure
`accuracy.hit_rate.hit_rate.HR`(…)	Hit Rate
`accuracy.map.map.MAP`(recommendations, …)	Mean Average Precision
`accuracy.mar.mar.MAR`(recommendations, …)	Mean Average Recall
`accuracy.mrr.mrr.MRR`(recommendations, …)	Mean Reciprocal Rank
`accuracy.ndcg.ndcg.NDCG`(recommendations, …)	normalized Discounted Cumulative Gain

Bias

`bias.aclt.aclt.ACLT`(recommendations, config, …)	Average coverage of long tail items
`bias.aplt.aplt.APLT`(recommendations, config, …)	Average percentage of long tail items
`bias.arp.arp.ARP`(recommendations, config, …)	Average Recommendation Popularity
`bias.pop_reo.pop_reo.PopREO`(recommendations, …)	Popularity-based Ranking-based Equal Opportunity
`bias.pop_reo.extended_pop_reo.ExtendedPopREO`(…)	Extended Popularity-based Ranking-based Equal Opportunity
`bias.pop_rsp.pop_rsp.PopRSP`(recommendations, …)	Popularity-based Ranking-based Statistical Parity
`bias.pop_rsp.extended_pop_rsp.ExtendedPopRSP`(…)	Extended Popularity-based Ranking-based Statistical Parity

Coverage

`coverage.item_coverage.item_coverage.ItemCoverage`(…)	Item Coverage
`coverage.num_retrieved.num_retrieved.NumRetrieved`(…)	Number of Recommendations Retrieved
`coverage.user_coverage.user_coverage.UserCoverage`(…)	User Coverage
`coverage.user_coverage.user_coverage_at_n.UserCoverageAtN`(…)	User Coverage on Top-N rec.

Diversity

`diversity.gini_index.gini_index.GiniIndex`(…)	Gini Index
`diversity.shannon_entropy.shannon_entropy.ShannonEntropy`(…)	Shannon Entropy
`diversity.SRecall.srecall.SRecall`(…)	Subtopic Recall

Fairness

`fairness.BiasDisparity.BiasDisparityBD`(…)	Bias Disparity - Standard
`fairness.BiasDisparity.BiasDisparityBR`(…)	Bias Disparity - Bias Recommendations
`fairness.BiasDisparity.BiasDisparityBS`(…)	Bias Disparity - Bias Source
`fairness.MAD.ItemMADranking.ItemMADranking`(…)	Item MAD Ranking-based
`fairness.MAD.ItemMADrating.ItemMADrating`(…)	Item MAD Rating-based
`fairness.MAD.UserMADranking.UserMADranking`(…)	User MAD Ranking-based
`fairness.MAD.UserMADrating.UserMADrating`(…)	User MAD Rating-based
`fairness.reo.reo.REO`(recommendations, …)	Ranking-based Equal Opportunity
`fairness.rsp.rsp.RSP`(recommendations, …)	Ranking-based Statistical Parity

Novelty

`novelty.EFD.efd.EFD`(recommendations, config, …)	Expected Free Discovery (EFD)
`novelty.EFD.extended_efd.ExtendedEFD`(…)	Extended EFD
`novelty.EPC.epc.EPC`(recommendations, config, …)	Expected Popularity Complement (EPC)
`novelty.EPC.extended_epc.ExtendedEPC`(…)	Extended EPC

Rating

`rating.mae.mae.MAE`(recommendations, config, …)	Mean Absolute Error
`rating.mse.mse.MSE`(recommendations, config, …)	Mean Squared Error
`rating.rmse.rmse.RMSE`(recommendations, …)	Root Mean Squared Error

Metrics Summary¶

Accuracy¶

Elliot integrates the following accuracy metrics.

Summary¶

`AUC.auc.AUC`(recommendations, config, params, …)	Area Under the Curve
`AUC.gauc.GAUC`(recommendations, config, …)	Group Area Under the Curve
`AUC.lauc.LAUC`(recommendations, config, …)	Limited Area Under the Curve
`DSC.dsc.DSC`(recommendations, config, params, …)	Sørensen–Dice coefficient
`f1.f1.F1`(recommendations, config, params, …)	F-Measure
`f1.extended_f1.ExtendedF1`(recommendations, …)	Extended F-Measure
`hit_rate.hit_rate.HR`(recommendations, …)	Hit Rate
`map.map.MAP`(recommendations, config, params, …)	Mean Average Precision
`mar.mar.MAR`(recommendations, config, params, …)	Mean Average Recall
`mrr.mrr.MRR`(recommendations, config, params, …)	Mean Reciprocal Rank
`ndcg.ndcg.NDCG`(recommendations, config, …)	normalized Discounted Cumulative Gain
`precision.precision.Precision`(…)	Precision-measure
`recall.recall.Recall`(recommendations, …)	Recall-measure

AUC¶

class elliot.evaluation.metrics.accuracy.AUC.auc.AUC(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Area Under the Curve

This class represents the implementation of the global AUC recommendation metric. Passing ‘AUC’ to the metrics list will enable the computation of the metric.

For further details, please refer to the AUC

Note

This metric does not calculate group-based AUC which considers the AUC scores averaged across users. It is also not limited to k. Instead, it calculates the scores on the entire prediction results regardless the users.

\[\mathrm {AUC} = \frac{\sum\limits_{i=1}^M rank_{i} - \frac {{M} \times {(M+1)}}{2}} {{{M} \times {N}}}\]

\(M\) is the number of positive samples.

\(N\) is the number of negative samples.

\(rank_i\) is the ascending rank of the ith positive sample.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [AUC]

GAUC¶

class elliot.evaluation.metrics.accuracy.AUC.gauc.GAUC(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Group Area Under the Curve

This class represents the implementation of the GroupAUC recommendation metric. Passing ‘GAUC’ to the metrics list will enable the computation of the metric.

“Deep Interest Network for Click-Through Rate Prediction” KDD ‘18 by Zhou, et al.

For further details, please refer to the paper

Note

It calculates the AUC score of each user, and finally obtains GAUC by weighting the user AUC. It is also not limited to k. Due to our padding for scores_tensor in RankEvaluator with -np.inf, the padding value will influence the ranks of origin items. Therefore, we use descending sort here and make an identity transformation to the formula of AUC, which is shown in auc_ function. For readability, we didn’t do simplification in the code.

\[\mathrm {GAUC} = \frac {{{M} \times {(M+N+1)} - \frac{M \times (M+1)}{2}} - \sum\limits_{i=1}^M rank_{i}} {{M} \times {N}}\]

\(M\) is the number of positive samples.

\(N\) is the number of negative samples.

\(rank_i\) is the descending rank of the ith positive sample.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [GAUC]

LAUC¶

class elliot.evaluation.metrics.accuracy.AUC.lauc.LAUC(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Limited Area Under the Curve

This class represents the implementation of the Limited AUC recommendation metric. Passing ‘LAUC’ to the metrics list will enable the computation of the metric.

“Setting Goals and Choosing Metrics for Recommender System Evaluations” by Gunnar Schröder, et al.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [LAUC]

DSC¶

class elliot.evaluation.metrics.accuracy.DSC.dsc.DSC(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Sørensen–Dice coefficient

This class represents the implementation of the Sørensen–Dice coefficient recommendation metric. Passing ‘DSC’ to the metrics list will enable the computation of the metric.

For further details, please refer to the page

\[\mathrm {DSC@K} = \frac{1+\beta^{2}}{\frac{1}{\text { metric_0@k }}+\frac{\beta^{2}}{\text { metric_1@k }}}\]

Parameters

beta – the beta coefficient (default: 1)
metric_0 – First considered metric (default: Precision)
metric_1 – Second considered metric (default: Recall)

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: DSC
  beta: 1
  metric_0: Precision
  metric_1: Recall

F1¶

class elliot.evaluation.metrics.accuracy.f1.f1.F1(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

F-Measure

This class represents the implementation of the F-score recommendation metric. Passing ‘F1’ to the metrics list will enable the computation of the metric.

For further details, please refer to the paper

\[\mathrm {F1@K} = \frac{1+\beta^{2}}{\frac{1}{\text { precision@k }}+\frac{\beta^{2}}{\text { recall@k }}}\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [F1]

Extended F1¶

class elliot.evaluation.metrics.accuracy.f1.extended_f1.ExtendedF1(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended F-Measure

This class represents the implementation of the F-score recommendation metric. Passing ‘ExtendedF1’ to the metrics list will enable the computation of the metric.

“Evaluating Recommender Systems” Gunawardana, Asela and Shani, Guy, In Recommender systems handbook pages 265–308, 2015

For further details, please refer to the paper

\[\mathrm {ExtendedF1@K} =\frac{2}{\frac{1}{\text { metric_0@k }}+\frac{1}{\text { metric_1@k }}}\]

Parameters

metric_0 – First considered metric (default: Precision)
metric_1 – Second considered metric (default: Recall)

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedF1
  metric_0: Precision
  metric_1: Recall

HR¶

class elliot.evaluation.metrics.accuracy.hit_rate.hit_rate.HR(recommendations: Dict[int, List[Tuple[int, float]]], config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Hit Rate

This class represents the implementation of the Hit Rate recommendation metric. Passing ‘HR’ to the metrics list will enable the computation of the metric.

For further details, please refer to the link

\[\mathrm {HR@K} =\frac{Number \space of \space Hits @K}{|GT|}\]

\(HR\) is the number of users with a positive sample in the recommendation list.

\(GT\) is the total number of samples in the test set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [HR]

MAP¶

class elliot.evaluation.metrics.accuracy.map.map.MAP(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Average Precision

This class represents the implementation of the Mean Average Precision recommendation metric. Passing ‘MAP’ to the metrics list will enable the computation of the metric.

For further details, please refer to the link

Note

In this case the normalization factor used is \(\frac{1}{\min (m,N)}\), which prevents your AP score from being unfairly suppressed when your number of recommendations couldn’t possibly capture all the correct ones.

\[\begin{split}\begin{align*} \mathrm{AP@N} &= \frac{1}{\mathrm{min}(m,N)}\sum_{k=1}^N P(k) \cdot rel(k) \\ \mathrm{MAP@N}& = \frac{1}{|U|}\sum_{u=1}^{|U|}(\mathrm{AP@N})_u \end{align*}\end{split}\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MAP]

MAR¶

class elliot.evaluation.metrics.accuracy.mar.mar.MAR(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Average Recall

This class represents the implementation of the Mean Average Recall recommendation metric. Passing ‘MAR’ to the metrics list will enable the computation of the metric.

For further details, please refer to the link

\[\begin{split}\begin{align*} \mathrm{Recall@N} &= \frac{1}{\mathrm{min}(m,|rel(k)|)}\sum_{k=1}^N P(k) \cdot rel(k) \\ \mathrm{MAR@N}& = \frac{1}{|U|}\sum_{u=1}^{|U|}(\mathrm{Recall@N})_u \end{align*}\end{split}\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MAR]

MRR¶

class elliot.evaluation.metrics.accuracy.mrr.mrr.MRR(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Reciprocal Rank

This class represents the implementation of the Mean Reciprocal Rank recommendation metric. Passing ‘MRR’ to the metrics list will enable the computation of the metric.

For further details, please refer to the link

\[\mathrm {MRR} = \frac{1}{|{U}|} \sum_{i=1}^{|{U}|} \frac{1}{rank_i}\]

\(U\) is the number of users, \(rank_i\) is the rank of the first item in the recommendation list in the test set results for user \(i\).

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MRR]

nDCG¶

class elliot.evaluation.metrics.accuracy.ndcg.ndcg.NDCG(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

normalized Discounted Cumulative Gain

This class represents the implementation of the nDCG recommendation metric.

For further details, please refer to the link

\[\begin{split}\begin{gather} \mathrm {DCG@K}=\sum_{i=1}^{K} \frac{2^{rel_i}-1}{\log_{2}{(i+1)}}\\ \mathrm {IDCG@K}=\sum_{i=1}^{K}\frac{1}{\log_{2}{(i+1)}}\\ \mathrm {NDCG_u@K}=\frac{DCG_u@K}{IDCG_u@K}\\ \mathrm {NDCG@K}=\frac{\sum \nolimits_{u \in u^{te}NDCG_u@K}}{|u^{te}|} \end{gather}\end{split}\]

\(K\) stands for recommending \(K\) items.

And the \(rel_i\) is the relevance of the item in position \(i\) in the recommendation list.

\(2^{rel_i}\) equals to 1 if the item hits otherwise 0.

\(U^{te}\) is for all users in the test set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [nDCG]

Precision¶

class elliot.evaluation.metrics.accuracy.precision.precision.Precision(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Precision-measure

This class represents the implementation of the Precision recommendation metric.

For further details, please refer to the link

\[\mathrm {Precision@K} = \frac{|Rel_u \cap Rec_u|}{Rec_u}\]

\(Rel_u\) is the set of items relevant to user \(U\),

\(Rec_u\) is the top K items recommended to users.

We obtain the result by calculating the average \(Precision@K\) of each user.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [Precision]

Recall¶

class elliot.evaluation.metrics.accuracy.recall.recall.Recall(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Recall-measure

This class represents the implementation of the Recall recommendation metric.

For further details, please refer to the link

\[\mathrm {Recall@K} = \frac{|Rel_u\cap Rec_u|}{Rel_u}\]

\(Rel_u\) is the set of items relevant to user \(U\),

\(Rec_u\) is the top K items recommended to users.

We obtain the result by calculating the average \(Recall@K\) of each user.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [Recall]

Rating¶

Elliot integrates the following ratings-based error metrics.

Summary¶

`mae.mae.MAE`(recommendations, config, params, …)	Mean Absolute Error
`mse.mse.MSE`(recommendations, config, params, …)	Mean Squared Error
`rmse.rmse.RMSE`(recommendations, config, …)	Root Mean Squared Error

MAE¶

class elliot.evaluation.metrics.rating.mae.mae.MAE(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Absolute Error

This class represents the implementation of the Mean Absolute Error recommendation metric.

For further details, please refer to the link

\[\mathrm{MAE}=\frac{1}{|{T}|} \sum_{(u, i) \in {T}}\left|\hat{r}_{u i}-r_{u i}\right|\]

\(T\) is the test set, \(\hat{r}_{u i}\) is the score predicted by the model,

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MAE]

MSE¶

class elliot.evaluation.metrics.rating.mse.mse.MSE(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Squared Error

This class represents the implementation of the Mean Squared Error recommendation metric.

For further details, please refer to the link

\[\mathrm{MSE} = \frac{1}{|{T}|} \sum_{(u, i) \in {T}}(\hat{r}_{u i}-r_{u i})^{2}\]

\(T\) is the test set, \(\hat{r}_{u i}\) is the score predicted by the model

\(r_{u i}\) the actual score of the test set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MSE]

RMSE¶

class elliot.evaluation.metrics.rating.rmse.rmse.RMSE(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Root Mean Squared Error

This class represents the implementation of the Root Mean Squared Error recommendation metric.

For further details, please refer to the link

\[\mathrm{RMSE} = \sqrt{\frac{1}{|{T}|} \sum_{(u, i) \in {T}}(\hat{r}_{u i}-r_{u i})^{2}}\]

\(T\) is the test set, \(\hat{r}_{u i}\) is the score predicted by the model

\(r_{u i}\) the actual score of the test set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [RMSE]

Coverage¶

Elliot integrates the following coverage metrics.

Summary¶

`item_coverage.item_coverage.ItemCoverage`(…)	Item Coverage
`num_retrieved.num_retrieved.NumRetrieved`(…)	Number of Recommendations Retrieved
`user_coverage.user_coverage.UserCoverage`(…)	User Coverage
`user_coverage.user_coverage_at_n.UserCoverageAtN`(…)	User Coverage on Top-N rec.

Item Coverage¶

class elliot.evaluation.metrics.coverage.item_coverage.item_coverage.ItemCoverage(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Item Coverage

This class represents the implementation of the Item Coverage recommendation metric.

For further details, please refer to the book

Note

The simplest measure of catalog coverage is the percentage of all items that can ever be recommended. This measure can be computed in many cases directly given the algorithm and the input data set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [ItemCoverage]

Number of Recommendations Retrieved¶

class elliot.evaluation.metrics.coverage.num_retrieved.num_retrieved.NumRetrieved(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Number of Recommendations Retrieved

This class represents the implementation of the Number of Recommendations Retrieved recommendation metric.

For further details, please refer to the link

simple_metrics: [NumRetrieved]

User Coverage¶

class elliot.evaluation.metrics.coverage.user_coverage.user_coverage.UserCoverage(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

User Coverage

This class represents the implementation of the User Coverage recommendation metric.

For further details, please refer to the book

Note

The proportion of users or user interactions for which the system can recommend items. In many applications the recommender may not provide recommendations for some users due to, e.g. low confidence in the accuracy of predictions for that user.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [UserCoverage]

User Coverage At N¶

class elliot.evaluation.metrics.coverage.user_coverage.user_coverage_at_n.UserCoverageAtN(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

User Coverage on Top-N rec. Lists

This class represents the implementation of the User Coverage recommendation metric.

For further details, please refer to the book

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [UserCoverageAtN]

Novelty¶

Elliot integrates the following novelty metrics.

Summary¶

`EFD.efd.EFD`(recommendations, config, params, …)	Expected Free Discovery (EFD)
`EFD.extended_efd.ExtendedEFD`(…)	Extended EFD
`EPC.epc.EPC`(recommendations, config, params, …)	Expected Popularity Complement (EPC)
`EPC.extended_epc.ExtendedEPC`(…)	Extended EPC

EFD¶

class elliot.evaluation.metrics.novelty.EFD.efd.EFD(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Expected Free Discovery (EFD)

This class represents the implementation of the Expected Free Discovery recommendation metric.

For further details, please refer to the paper

Note

EFD can be read as the expected ICF of seen recommended items

\[\mathrm {EFD}=C \sum_{i_{k} \in R} {disc}(k) p({rel} \mid i_{k}, u)( -\log _{2} p(i \mid {seen}, \theta))\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [EFD]

Extended EFD¶

class elliot.evaluation.metrics.novelty.EFD.extended_efd.ExtendedEFD(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended EFD

This class represents the implementation of the Extended Expected Free Discovery recommendation metric.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedEFD

EPC¶

class elliot.evaluation.metrics.novelty.EPC.epc.EPC(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Expected Popularity Complement (EPC)

This class represents the implementation of the Expected Popularity Complement recommendation metric.

For further details, please refer to the paper

Note

EPC can be read as the expected number of seen relevant recommended items not previously seen

\[\mathrm{EPC}=C \sum_{i_{k} \in R} \operatorname{disc}(k) p\left(r e l \mid i_{k}, u\right)\left(1-p\left(\operatorname{seen} \mid t_{k}\right)\right)\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [EPC]

Extended EPC¶

class elliot.evaluation.metrics.novelty.EPC.extended_epc.ExtendedEPC(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended EPC

This class represents the implementation of the Extended EPC recommendation metric.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedEPC

Diversity¶

Elliot integrates the following diversity metrics.

Summary¶

`gini_index.gini_index.GiniIndex`(…)	Gini Index
`shannon_entropy.shannon_entropy.ShannonEntropy`(…)	Shannon Entropy
`SRecall.srecall.SRecall`(recommendations, …)	Subtopic Recall

Gini Index¶

class elliot.evaluation.metrics.diversity.gini_index.gini_index.GiniIndex(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Gini Index

This class represents the implementation of the Gini Index recommendation metric.

For further details, please refer to the book

\[\mathrm {GiniIndex}=\frac{1}{n-1} \sum_{j=1}^{n}(2 j-n-1) p\left(i_{j}\right)\]

\(i_{j}\) is the list of items ordered according to increasing p(i)

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [Gini]

Shannon Entropy¶

class elliot.evaluation.metrics.diversity.shannon_entropy.shannon_entropy.ShannonEntropy(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Shannon Entropy

This class represents the implementation of the Shannon Entropy recommendation metric.

For further details, please refer to the book

\[\mathrm {ShannonEntropy}=-\sum_{i=1}^{n} p(i) \log p(i)\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [SEntropy]

SRecall¶

class elliot.evaluation.metrics.diversity.SRecall.srecall.SRecall(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Subtopic Recall

This class represents the implementation of the Subtopic Recall (S-Recall) recommendation metric.

For further details, please refer to the paper

\[\mathrm {SRecall}=\frac{\left|\cup_{i=1}^{K} {subtopics}\left(d_{i}\right)\right|}{n_{A}}\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [SRecall]

Bias¶

Elliot integrates the following bias metrics.

Summary¶

`aclt.aclt.ACLT`(recommendations, config, …)	Average coverage of long tail items
`aplt.aplt.APLT`(recommendations, config, …)	Average percentage of long tail items
`arp.arp.ARP`(recommendations, config, params, …)	Average Recommendation Popularity
`pop_reo.pop_reo.PopREO`(recommendations, …)	Popularity-based Ranking-based Equal Opportunity
`pop_reo.extended_pop_reo.ExtendedPopREO`(…)	Extended Popularity-based Ranking-based Equal Opportunity
`pop_rsp.pop_rsp.PopRSP`(recommendations, …)	Popularity-based Ranking-based Statistical Parity
`pop_rsp.extended_pop_rsp.ExtendedPopRSP`(…)	Extended Popularity-based Ranking-based Statistical Parity

ACLT¶

class elliot.evaluation.metrics.bias.aclt.aclt.ACLT(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Average coverage of long tail items

This class represents the implementation of the Average coverage of long tail items recommendation metric.

For further details, please refer to the paper

\[\mathrm {ACLT}=\frac{1}{\left|U_{t}\right|} \sum_{u \in U_{f}} \sum_{i \in L_{u}} 1(i \in \Gamma)\]

\(U_{t}\) is the number of users in the test set.

\(L_{u}\) is the recommended list of items for user u.

\(1(i \in \Gamma)\) is an indicator function and it equals to 1 when i is in Gamma.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [ACLT]

APLT¶

class elliot.evaluation.metrics.bias.aplt.aplt.APLT(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Average percentage of long tail items

This class represents the implementation of the Average percentage of long tail items recommendation metric.

For further details, please refer to the paper

\[\mathrm {ACLT}=\frac{1}{\left|U_{t}\right|} \sum_{u \in U_{t}} \frac{|\{i, i \in(L(u) \cap \sim \Phi)\}|}{|L(u)|}\]

\(U_{t}\) is the number of users in the test set.

\(L_{u}\) is the recommended list of items for user u.

\(\sim \Phi\) medium-tail items.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [APLT]

ARP¶

class elliot.evaluation.metrics.bias.arp.arp.ARP(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Average Recommendation Popularity

This class represents the implementation of the Average Recommendation Popularity recommendation metric.

For further details, please refer to the paper

\[\mathrm {ARP}=\frac{1}{\left|U_{t}\right|} \sum_{u \in U_{t}} \frac{\sum_{i \in L_{u}} \phi(i)}{\left|L_{u}\right|}\]

\(U_{t}\) is the number of users in the test set.

\(L_{u}\) is the recommended list of items for user u.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [ARP]

PopREO¶

class elliot.evaluation.metrics.bias.pop_reo.pop_reo.PopREO(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Popularity-based Ranking-based Equal Opportunity

This class represents the implementation of the Popularity-based Ranking-based Equal Opportunity (REO) recommendation metric.

For further details, please refer to the paper

\[\mathrm {REO}=\frac{{std}\left(P\left(R @ k \mid g=g_{1}, y=1\right) \ldots P\left(R(a) k=g_{A}, y=1\right)\right)} {{mean}\left(P\left(R @ k \mid g=g_{1}, y=1\right) \ldots P\left(R @ k \mid g=g_{A}, y=1\right)\right)}\]

\(P\left(R @ k \mid g=g_{a}, y=1\right) = \frac{\sum_{u=1}^{N} \sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right) Y\left(u, R_{u, i}\right)} {\sum_{u=1}^{N} \sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i) Y(u, i)}\)

\(Y\left(u, R_{u, i}\right)\) identifies the ground-truth label of a user-item pair left(u, R_{u, i}right), if item R_{u, i} is liked by user 𝑢, returns 1, otherwise 0

\(\sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right) Y\left(u, R_{u, i}\right)\) counts how many items in test set from group {g_a} are ranked in top-𝑘 for user u

\(\sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i) Y(u, i)\) counts the total number of items from group {g_a} 𝑎 in test set for user u

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [PopREO]

Extended PopREO¶

class elliot.evaluation.metrics.bias.pop_reo.extended_pop_reo.ExtendedPopREO(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended Popularity-based Ranking-based Equal Opportunity

This class represents the implementation of the Extended Popularity-based Ranking-based Equal Opportunity (REO) recommendation metric.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedPopREO

PopRSP¶

class elliot.evaluation.metrics.bias.pop_rsp.pop_rsp.PopRSP(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Popularity-based Ranking-based Statistical Parity

This class represents the implementation of the Popularity-based Ranking-based Statistical Parity (RSP) recommendation metric.

For further details, please refer to the paper

\[\mathrm {RSP}=\frac{{std}\left(P\left(R @ k \mid g=g_{1}\right), \ldots, P\left(R @ k \mid g=g_{A}\right)\right)} {{mean}\left(P\left(R @ k \mid g=g_{1}\right), \ldots, P\left(R @ k \mid g=g_{A}\right)\right)}\]

:math P(R @ k mid g=g_{A})) = frac{sum_{u=1}^{N} sum_{i=1}^{k} G_{g_{a}}(R_{u, i})} {sum_{u=1}^{N} sum_{i in I backslash I_{u}^{+}} G_{g_{a}}(i)}

\(\sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right)\) calculates how many un-interacted items from group {g_a} are ranked in top-𝑘 for user u.

\(\sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i)\) calculates how many un-interacted items belong to group {g_a} for u

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [PopRSP]

Extended PopRSP¶

class elliot.evaluation.metrics.bias.pop_rsp.extended_pop_rsp.ExtendedPopRSP(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended Popularity-based Ranking-based Statistical Parity

This class represents the implementation of the Extended Popularity-based Ranking-based Statistical Parity (RSP) recommendation metric.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedPopRSP

Fairness¶

Elliot integrates the following fairness metrics.

Summary¶

`BiasDisparity.BiasDisparityBD`(…)	Bias Disparity - Standard
`BiasDisparity.BiasDisparityBR`(…)	Bias Disparity - Bias Recommendations
`BiasDisparity.BiasDisparityBS`(…)	Bias Disparity - Bias Source
`MAD.ItemMADranking.ItemMADranking`(…)	Item MAD Ranking-based
`MAD.ItemMADrating.ItemMADrating`(…)	Item MAD Rating-based
`MAD.UserMADranking.UserMADranking`(…)	User MAD Ranking-based
`MAD.UserMADrating.UserMADrating`(…)	User MAD Rating-based
`reo.reo.REO`(recommendations, config, params, …)	Ranking-based Equal Opportunity
`rsp.rsp.RSP`(recommendations, config, params, …)	Ranking-based Statistical Parity

BiasDisparity BD¶

class elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBD.BiasDisparityBD(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Bias Disparity - Standard

This class represents the implementation of the Bias Disparity recommendation metric.

For further details, please refer to the paper

\[\mathrm {BD(G, C)}=\frac{B_{R}(G, C)-B_{S}(G, C)}{B_{S}(G, C)}\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
    - metric: BiasDisparityBD
      user_clustering_name: Happiness
      user_clustering_file: ../data/movielens_1m/u_happy.tsv
      item_clustering_name: ItemPopularity
      item_clustering_file: ../data/movielens_1m/i_pop.tsv

BiasDisparity BR¶

class elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBR.BiasDisparityBR(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Bias Disparity - Bias Recommendations

This class represents the implementation of the Bias Disparity - Bias Recommendations recommendation metric.

For further details, please refer to the paper

\[\mathrm {BD(G, C)}=\frac{B_{R}(G, C)-B_{S}(G, C)}{B_{S}(G, C)}\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
    - metric: BiasDisparityBR
      user_clustering_name: Happiness
      user_clustering_file: ../data/movielens_1m/u_happy.tsv
      item_clustering_name: ItemPopularity
      item_clustering_file: ../data/movielens_1m/i_pop.tsv

BiasDisparity BS¶

class elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBS.BiasDisparityBS(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Bias Disparity - Bias Source

This class represents the implementation of the Bias Disparity - Bias Source recommendation metric.

For further details, please refer to the paper

\[\mathrm {B_{S}(G, C)}=\frac{P R_{S}(G, C)}{P(C)}\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
    - metric: BiasDisparityBS
      user_clustering_name: Happiness
      user_clustering_file: ../data/movielens_1m/u_happy.tsv
      item_clustering_name: ItemPopularity
      item_clustering_file: ../data/movielens_1m/i_pop.tsv

ItemMADranking¶

class elliot.evaluation.metrics.fairness.MAD.ItemMADranking.ItemMADranking(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Item MAD Ranking-based

This class represents the implementation of the Item MAD ranking recommendation metric.

For further details, please refer to the paper

\[\mathrm {MAD}={avg}_{i, j}({MAD}(R^{(i)}, R^{(j)}))\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ItemMADranking
  clustering_name: ItemPopularity
  clustering_file: ../data/movielens_1m/i_pop.tsv

ItemMADrating¶

class elliot.evaluation.metrics.fairness.MAD.ItemMADrating.ItemMADrating(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Item MAD Rating-based

This class represents the implementation of the Item MAD rating recommendation metric.

For further details, please refer to the paper

\[\mathrm {MAD}={avg}_{i, j}({MAD}(R^{(i)}, R^{(j)}))\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ItemMADrating
  clustering_name: ItemPopularity
  clustering_file: ../data/movielens_1m/i_pop.tsv

UserMADranking¶

class elliot.evaluation.metrics.fairness.MAD.UserMADranking.UserMADranking(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

User MAD Ranking-based

This class represents the implementation of the User MAD ranking recommendation metric.

For further details, please refer to the paper

\[\mathrm {MAD}={avg}_{i, j}({MAD}(R^{(i)}, R^{(j)}))\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: UserMADranking
  clustering_name: Happiness
  clustering_file: ../data/movielens_1m/u_happy.tsv

UserMADrating¶

class elliot.evaluation.metrics.fairness.MAD.UserMADrating.UserMADrating(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

User MAD Rating-based

This class represents the implementation of the User MAD rating recommendation metric.

For further details, please refer to the paper

\[\mathrm {MAD}={avg}_{i, j}({MAD}(R^{(i)}, R^{(j)}))\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: UserMADrating
  clustering_name: Happiness
  clustering_file: ../data/movielens_1m/u_happy.tsv

REO¶

class elliot.evaluation.metrics.fairness.reo.reo.REO(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Ranking-based Equal Opportunity

This class represents the implementation of the Ranking-based Equal Opportunity (REO) recommendation metric.

For further details, please refer to the paper

\[\mathrm {REO}=\frac{{std}\left(P\left(R @ k \mid g=g_{1}, y=1\right) \ldots P\left(R(a) k=g_{A}, y=1\right)\right)} {{mean}\left(P\left(R @ k \mid g=g_{1}, y=1\right) \ldots P\left(R @ k \mid g=g_{A}, y=1\right)\right)}\]

\(P\left(R @ k \mid g=g_{a}, y=1\right) = \frac{\sum_{u=1}^{N} \sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right) Y\left(u, R_{u, i}\right)} {\sum_{u=1}^{N} \sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i) Y(u, i)}\)

\(Y\left(u, R_{u, i}\right)\) identifies the ground-truth label of a user-item pair left(u, R_{u, i}right), if item R_{u, i} is liked by user 𝑢, returns 1, otherwise 0

\(\sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right) Y\left(u, R_{u, i}\right)\) counts how many items in test set from group {g_a} are ranked in top-𝑘 for user u

\(\sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i) Y(u, i)\) counts the total number of items from group {g_a} 𝑎 in test set for user u

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
 - metric: REO
  clustering_name: ItemPopularity
  clustering_file: ../data/movielens_1m/i_pop.tsv

RSP¶

class elliot.evaluation.metrics.fairness.rsp.rsp.RSP(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Ranking-based Statistical Parity

This class represents the implementation of the Ranking-based Statistical Parity (RSP) recommendation metric.

For further details, please refer to the paper

\[\mathrm {RSP}=\frac{{std}(P(R @ k \mid g=g_{1}), \ldots, P(R @ k \mid g=g_{A}))} {{mean}(P(R @ k \mid g=g_{1}), \ldots, P(R @ k \mid g=g_{A}))}\]

\(P(R @ k \mid g=g_{A})) = \frac{\sum_{u=1}^{N} \sum_{i=1}^{k} G_{g_{a}}(R_{u, i})} {\sum_{u=1}^{N} \sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i)}\)

\(\sum_{i=1}^{k} G_{g_{a}}(R_{u, i})\) calculates how many un-interacted items from group {g_a} are ranked in top-𝑘 for user u.

\(\sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i)\) calculates how many un-interacted items belong to group {g_a} for u

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
 - metric: RSP
  clustering_name: ItemPopularity
  clustering_file: ../data/movielens_1m/i_pop.tsv

elliot package¶

Subpackages¶

elliot.dataset package¶

Subpackages¶

elliot.dataset.dataloader package¶

Submodules¶

elliot.dataset.dataloader.knowledge_aware_chains module¶

Module description:

class elliot.dataset.dataloader.knowledge_aware_chains.KnowledgeChainsDataObject(config, data_tuple, side_information_data, *args, **kwargs)[source]¶

Bases: object

Load train and test dataset

build_dict(dataframe, users)[source]¶

build_sparse()[source]¶

build_sparse_ratings()[source]¶

dataframe_to_dict(data)[source]¶

get_test()[source]¶

get_validation()[source]¶

class elliot.dataset.dataloader.knowledge_aware_chains.KnowledgeChainsLoader(config, *args, **kwargs)[source]¶

Bases: object

Load train and test dataset

check_timestamp(d: pandas.DataFrame) → pandas.DataFrame[source]¶

generate_dataobjects() → List[object][source]¶

generate_dataobjects_mock() → List[object][source]¶

load_attribute_file(attribute_file, separator='\t')[source]¶

load_dataset_dataframe(file_ratings, separator='\t', attribute_file=None, feature_file=None, properties_file=None, column_names=['userId', 'itemId', 'rating', 'timestamp'], additive=True, threshold=10)[source]¶

load_dataset_dict(file_ratings, separator='\t', attribute_file=None, feature_file=None, properties_file=None, additive=True, threshold=10)[source]¶

load_feature_names(infile, separator='\t')[source]¶

load_item_set(ratings_file, separator='\t', itemPosition=1)[source]¶

load_properties(properties_file)[source]¶

read_splitting(folder_path)[source]¶

reduce_attribute_map_property_selection(map, items, feature_names, properties, additive, threshold=10)[source]¶

reduce_dataset_by_item_list(ratings_file, items, separator='\t')[source]¶

elliot.dataset.dataloader.visual_dataloader module¶

Module description:

class elliot.dataset.dataloader.visual_dataloader.VisualDataObject(config, data_tuple, side_information_data, *args, **kwargs)[source]¶

Bases: object

Load train and test dataset

build_dict(dataframe, users)[source]¶

build_sparse()[source]¶

build_sparse_ratings()[source]¶

dataframe_to_dict(data)[source]¶

get_test()[source]¶

get_validation()[source]¶

read_images(images_folder, image_set, size_tuple)[source]¶

read_images_multiprocessing(images_folder, image_set, size_tuple)[source]¶

static read_single_image(images_folder, image_set, size_tuple, image_path)[source]¶

class elliot.dataset.dataloader.visual_dataloader.VisualLoader(config, *args, **kwargs)[source]¶

Bases: object

Load train and test dataset

check_timestamp(d: pandas.DataFrame) → pandas.DataFrame[source]¶

generate_dataobjects() → List[object][source]¶

generate_dataobjects_mock() → List[object][source]¶

load_dataset_dataframe(file_ratings, separator='\t', visual_feature_set=None, column_names=['userId', 'itemId', 'rating', 'timestamp'])[source]¶

read_splitting(folder_path)[source]¶

reduce_dataset_by_item_list(ratings_file, items, separator='\t')[source]¶

Module contents¶

elliot.dataset.samplers package¶

Submodules¶

elliot.dataset.samplers.custom_pointwise_sparse_sampler module¶

Module description:

class elliot.dataset.samplers.custom_pointwise_sparse_sampler.Sampler(indexed_ratings, sp_i_train)[source]¶

Bases: object

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.custom_sampler module¶

Module description:

class elliot.dataset.samplers.custom_sampler.Sampler(indexed_ratings)[source]¶

Bases: object

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.custom_sparse_sampler module¶

Module description:

class elliot.dataset.samplers.custom_sparse_sampler.Sampler(indexed_ratings, sp_i_train)[source]¶

Bases: object

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.pairwise_sampler module¶

Module description:

class elliot.dataset.samplers.pairwise_sampler.Sampler(ratings, users, items)[source]¶

Bases: object

step(events: int)[source]¶

elliot.dataset.samplers.pipeline_sampler module¶

Module description:

class elliot.dataset.samplers.pipeline_sampler.Sampler(indexed_ratings, item_indices, images_path, output_image_size, epochs)[source]¶

Bases: object

pipeline(num_users, batch_size)[source]¶

pipeline_eval(batch_size)[source]¶

read_image(item)[source]¶

read_images_triple(user, pos, neg)[source]¶

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.pointwise_cfgan_sampler module¶

Module description:

class elliot.dataset.samplers.pointwise_cfgan_sampler.Sampler(indexed_ratings, sp_i_train, s_zr, s_pm)[source]¶

Bases: object

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.pointwise_pos_neg_ratings_sampler module¶

Module description:

class elliot.dataset.samplers.pointwise_pos_neg_ratings_sampler.Sampler(indexed_ratings, sparse_i_ratings)[source]¶

Bases: object

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.pointwise_pos_neg_ratio_ratings_sampler module¶

Module description:

class elliot.dataset.samplers.pointwise_pos_neg_ratio_ratings_sampler.Sampler(indexed_ratings, sparse_i_ratings, neg_ratio)[source]¶

Bases: object

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.pointwise_pos_neg_sampler module¶

Module description:

class elliot.dataset.samplers.pointwise_pos_neg_sampler.Sampler(indexed_ratings)[source]¶

Bases: object

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.pointwise_wide_and_deep_sampler module¶

Module description:

class elliot.dataset.samplers.pointwise_wide_and_deep_sampler.Sampler(data)[source]¶

Bases: object

step(events: int, batch_size: int)[source]¶

elliot.dataset.samplers.sparse_sampler module¶

Module description:

class elliot.dataset.samplers.sparse_sampler.Sampler(sp_i_train)[source]¶

Bases: object

step(users: int, batch_size: int)[source]¶

Module contents¶

Module description:

Submodules¶

elliot.dataset.abstract_dataset module¶

class elliot.dataset.abstract_dataset.AbstractDataset(*args, **kwargs)[source]¶

Bases: object

abstract build_dict()[source]¶

abstract build_sparse(*args)[source]¶

abstract get_test(*args)[source]¶

required_attributes = ['config', 'args', 'kwargs', 'users', 'items', 'num_users', 'num_items', 'private_users', 'public_users', 'private_items', 'public_items', 'transactions', 'train_dict', 'i_train_dict', 'sp_i_train', 'test_dict']¶

class elliot.dataset.abstract_dataset.ForceRequiredAttributeDefinitionMeta[source]¶

Bases: type

check_required_attributes(class_object)[source]¶

elliot.dataset.dataset module¶

Module description:

class elliot.dataset.dataset.DataSet(*args, **kwargs)[source]¶

Bases: elliot.dataset.abstract_dataset.AbstractDataset

Load train and test dataset

build_dict(dataframe, users)[source]¶

build_sparse()[source]¶

build_sparse_ratings()[source]¶

dataframe_to_dict(data)[source]¶

get_test()[source]¶

get_validation()[source]¶

class elliot.dataset.dataset.DataSetLoader(config, *args, **kwargs)[source]¶

Bases: object

Load train and test dataset

check_timestamp(d: pandas.DataFrame) → pandas.DataFrame[source]¶

generate_dataobjects() → List[object][source]¶

generate_dataobjects_mock() → List[object][source]¶

read_splitting(folder_path)[source]¶

Module contents¶

Module description:

elliot.evaluation package¶

Subpackages¶

elliot.evaluation.metrics package¶

Subpackages¶

elliot.evaluation.metrics.accuracy package¶

Subpackages¶

elliot.evaluation.metrics.accuracy.AUC package¶

Submodules¶

elliot.evaluation.metrics.accuracy.AUC.auc module¶

This is the implementation of the global AUC metric. It proceeds from a system-wise computation.

class elliot.evaluation.metrics.accuracy.AUC.auc.AUC(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Area Under the Curve

This class represents the implementation of the global AUC recommendation metric. Passing ‘AUC’ to the metrics list will enable the computation of the metric.

For further details, please refer to the AUC

Note

This metric does not calculate group-based AUC which considers the AUC scores averaged across users. It is also not limited to k. Instead, it calculates the scores on the entire prediction results regardless the users.

\[\mathrm {AUC} = \frac{\sum\limits_{i=1}^M rank_{i} - \frac {{M} \times {(M+1)}}{2}} {{{M} \times {N}}}\]

\(M\) is the number of positive samples.

\(N\) is the number of negative samples.

\(rank_i\) is the ascending rank of the ith positive sample.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [AUC]

eval()[source]¶: Evaluation function :return: the overall value of AUC

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

static needs_full_recommendations()[source]¶

elliot.evaluation.metrics.accuracy.AUC.gauc module¶

This is the implementation of the GroupAUC metric. It proceeds from a user-wise computation, and average the AUC values over the users.

class elliot.evaluation.metrics.accuracy.AUC.gauc.GAUC(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Group Area Under the Curve

This class represents the implementation of the GroupAUC recommendation metric. Passing ‘GAUC’ to the metrics list will enable the computation of the metric.

“Deep Interest Network for Click-Through Rate Prediction” KDD ‘18 by Zhou, et al.

For further details, please refer to the paper

Note

It calculates the AUC score of each user, and finally obtains GAUC by weighting the user AUC. It is also not limited to k. Due to our padding for scores_tensor in RankEvaluator with -np.inf, the padding value will influence the ranks of origin items. Therefore, we use descending sort here and make an identity transformation to the formula of AUC, which is shown in auc_ function. For readability, we didn’t do simplification in the code.

\[\mathrm {GAUC} = \frac {{{M} \times {(M+N+1)} - \frac{M \times (M+1)}{2}} - \sum\limits_{i=1}^M rank_{i}} {{M} \times {N}}\]

\(M\) is the number of positive samples.

\(N\) is the number of negative samples.

\(rank_i\) is the descending rank of the ith positive sample.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [GAUC]

eval()[source]¶: Evaluation function :return: the overall averaged value of AUC

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of AUC per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

static needs_full_recommendations()[source]¶

elliot.evaluation.metrics.accuracy.AUC.lauc module¶

This is the implementation of the Limited AUC metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.AUC.lauc.LAUC(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Limited Area Under the Curve

This class represents the implementation of the Limited AUC recommendation metric. Passing ‘LAUC’ to the metrics list will enable the computation of the metric.

“Setting Goals and Choosing Metrics for Recommender System Evaluations” by Gunnar Schröder, et al.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [LAUC]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of LAUC per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.accuracy.DSC package¶

Submodules¶

elliot.evaluation.metrics.accuracy.DSC.dsc module¶

This is the implementation of the Sørensen–Dice coefficient metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.DSC.dsc.DSC(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Sørensen–Dice coefficient

This class represents the implementation of the Sørensen–Dice coefficient recommendation metric. Passing ‘DSC’ to the metrics list will enable the computation of the metric.

For further details, please refer to the page

\[\mathrm {DSC@K} = \frac{1+\beta^{2}}{\frac{1}{\text { metric_0@k }}+\frac{\beta^{2}}{\text { metric_1@k }}}\]

Parameters

beta – the beta coefficient (default: 1)
metric_0 – First considered metric (default: Precision)
metric_1 – Second considered metric (default: Recall)

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: DSC
  beta: 1
  metric_0: Precision
  metric_1: Recall

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Sørensen–Dice coefficient per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.accuracy.f1 package¶

Submodules¶

elliot.evaluation.metrics.accuracy.f1.extended_f1 module¶

This is the implementation of the F-score metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.f1.extended_f1.ExtendedF1(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended F-Measure

This class represents the implementation of the F-score recommendation metric. Passing ‘ExtendedF1’ to the metrics list will enable the computation of the metric.

“Evaluating Recommender Systems” Gunawardana, Asela and Shani, Guy, In Recommender systems handbook pages 265–308, 2015

For further details, please refer to the paper

\[\mathrm {ExtendedF1@K} =\frac{2}{\frac{1}{\text { metric_0@k }}+\frac{1}{\text { metric_1@k }}}\]

Parameters

metric_0 – First considered metric (default: Precision)
metric_1 – Second considered metric (default: Recall)

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedF1
  metric_0: Precision
  metric_1: Recall

eval_user_metric()[source]¶

get()[source]¶

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

process()[source]¶: Evaluation function :return: the overall value of Bias Disparity

elliot.evaluation.metrics.accuracy.f1.f1 module¶

This is the implementation of the F-score metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.f1.f1.F1(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

F-Measure

This class represents the implementation of the F-score recommendation metric. Passing ‘F1’ to the metrics list will enable the computation of the metric.

For further details, please refer to the paper

\[\mathrm {F1@K} = \frac{1+\beta^{2}}{\frac{1}{\text { precision@k }}+\frac{\beta^{2}}{\text { recall@k }}}\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [F1]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of F-score

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.accuracy.hit_rate package¶

Submodules¶

elliot.evaluation.metrics.accuracy.hit_rate.hit_rate module¶

This is the implementation of the Hit Rate metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.hit_rate.hit_rate.HR(recommendations: Dict[int, List[Tuple[int, float]]], config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Hit Rate

This class represents the implementation of the Hit Rate recommendation metric. Passing ‘HR’ to the metrics list will enable the computation of the metric.

For further details, please refer to the link

\[\mathrm {HR@K} =\frac{Number \space of \space Hits @K}{|GT|}\]

\(HR\) is the number of users with a positive sample in the recommendation list.

\(GT\) is the total number of samples in the test set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [HR]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Hit Rate per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.accuracy.map package¶

Submodules¶

elliot.evaluation.metrics.accuracy.map.map module¶

This is the implementation of the Mean Average Precision metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.map.map.MAP(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Average Precision

This class represents the implementation of the Mean Average Precision recommendation metric. Passing ‘MAP’ to the metrics list will enable the computation of the metric.

For further details, please refer to the link

Note

In this case the normalization factor used is \(\frac{1}{\min (m,N)}\), which prevents your AP score from being unfairly suppressed when your number of recommendations couldn’t possibly capture all the correct ones.

\[\begin{split}\begin{align*} \mathrm{AP@N} &= \frac{1}{\mathrm{min}(m,N)}\sum_{k=1}^N P(k) \cdot rel(k) \\ \mathrm{MAP@N}& = \frac{1}{|U|}\sum_{u=1}^{|U|}(\mathrm{AP@N})_u \end{align*}\end{split}\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MAP]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Mean Average Precision per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.accuracy.mar package¶

Submodules¶

elliot.evaluation.metrics.accuracy.mar.mar module¶

This is the implementation of the Mean Average Recall metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.mar.mar.MAR(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Average Recall

This class represents the implementation of the Mean Average Recall recommendation metric. Passing ‘MAR’ to the metrics list will enable the computation of the metric.

For further details, please refer to the link

\[\begin{split}\begin{align*} \mathrm{Recall@N} &= \frac{1}{\mathrm{min}(m,|rel(k)|)}\sum_{k=1}^N P(k) \cdot rel(k) \\ \mathrm{MAR@N}& = \frac{1}{|U|}\sum_{u=1}^{|U|}(\mathrm{Recall@N})_u \end{align*}\end{split}\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MAR]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Mean Average Recall per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.accuracy.mrr package¶

Submodules¶

elliot.evaluation.metrics.accuracy.mrr.mrr module¶

This is the implementation of the Mean Reciprocal Rank metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.mrr.mrr.MRR(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Reciprocal Rank

This class represents the implementation of the Mean Reciprocal Rank recommendation metric. Passing ‘MRR’ to the metrics list will enable the computation of the metric.

For further details, please refer to the link

\[\mathrm {MRR} = \frac{1}{|{U}|} \sum_{i=1}^{|{U}|} \frac{1}{rank_i}\]

\(U\) is the number of users, \(rank_i\) is the rank of the first item in the recommendation list in the test set results for user \(i\).

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MRR]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Mean Reciprocal Rank per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.accuracy.ndcg package¶

Submodules¶

elliot.evaluation.metrics.accuracy.ndcg.ndcg module¶

This is the implementation of the normalized Discounted Cumulative Gain metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.ndcg.ndcg.NDCG(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

normalized Discounted Cumulative Gain

This class represents the implementation of the nDCG recommendation metric.

For further details, please refer to the link

\[\begin{split}\begin{gather} \mathrm {DCG@K}=\sum_{i=1}^{K} \frac{2^{rel_i}-1}{\log_{2}{(i+1)}}\\ \mathrm {IDCG@K}=\sum_{i=1}^{K}\frac{1}{\log_{2}{(i+1)}}\\ \mathrm {NDCG_u@K}=\frac{DCG_u@K}{IDCG_u@K}\\ \mathrm {NDCG@K}=\frac{\sum \nolimits_{u \in u^{te}NDCG_u@K}}{|u^{te}|} \end{gather}\end{split}\]

\(K\) stands for recommending \(K\) items.

And the \(rel_i\) is the relevance of the item in position \(i\) in the recommendation list.

\(2^{rel_i}\) equals to 1 if the item hits otherwise 0.

\(U^{te}\) is for all users in the test set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [nDCG]

compute_idcg(user, cutoff: int) → float[source]¶: Method to compute Ideal Discounted Cumulative Gain :param gain_map: :param cutoff: :return:

compute_user_ndcg(user_recommendations: List, user, cutoff: int) → float[source]¶: Method to compute normalized Discounted Cumulative Gain :param sorted_item_predictions: :param gain_map: :param cutoff: :return:

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of normalized Discounted Cumulative Gain per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

This is the nDCG metric module.

This module contains and expose the recommendation metric.

elliot.evaluation.metrics.accuracy.precision package¶

Submodules¶

elliot.evaluation.metrics.accuracy.precision.precision module¶

This is the implementation of the Precision metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.precision.precision.Precision(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Precision-measure

This class represents the implementation of the Precision recommendation metric.

For further details, please refer to the link

\[\mathrm {Precision@K} = \frac{|Rel_u \cap Rec_u|}{Rec_u}\]

\(Rel_u\) is the set of items relevant to user \(U\),

\(Rec_u\) is the top K items recommended to users.

We obtain the result by calculating the average \(Precision@K\) of each user.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [Precision]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Precision

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

This is the Precision metric module.

This module contains and expose the recommendation metric.

elliot.evaluation.metrics.accuracy.recall package¶

Submodules¶

elliot.evaluation.metrics.accuracy.recall.recall module¶

This is the implementation of the Recall metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.accuracy.recall.recall.Recall(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Recall-measure

This class represents the implementation of the Recall recommendation metric.

For further details, please refer to the link

\[\mathrm {Recall@K} = \frac{|Rel_u\cap Rec_u|}{Rel_u}\]

\(Rel_u\) is the set of items relevant to user \(U\),

\(Rec_u\) is the top K items recommended to users.

We obtain the result by calculating the average \(Recall@K\) of each user.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [Recall]

eval_user_metric()[source]¶: Evaluation Function :return: the overall averaged value of Recall per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

This is the Recall metric implementation.

This module contains and expose the recommendation metric.

Module contents¶

elliot.evaluation.metrics.bias package¶

Subpackages¶

elliot.evaluation.metrics.bias.aclt package¶

Submodules¶

elliot.evaluation.metrics.bias.aclt.aclt module¶

This is the implementation of the Average coverage of long tail items metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.bias.aclt.aclt.ACLT(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Average coverage of long tail items

This class represents the implementation of the Average coverage of long tail items recommendation metric.

For further details, please refer to the paper

\[\mathrm {ACLT}=\frac{1}{\left|U_{t}\right|} \sum_{u \in U_{f}} \sum_{i \in L_{u}} 1(i \in \Gamma)\]

\(U_{t}\) is the number of users in the test set.

\(L_{u}\) is the recommended list of items for user u.

\(1(i \in \Gamma)\) is an indicator function and it equals to 1 when i is in Gamma.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [ACLT]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of ACLT

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.bias.aplt package¶

Submodules¶

elliot.evaluation.metrics.bias.aplt.aplt module¶

This is the implementation of the Average percentage of long tail items metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.bias.aplt.aplt.APLT(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Average percentage of long tail items

This class represents the implementation of the Average percentage of long tail items recommendation metric.

For further details, please refer to the paper

\[\mathrm {ACLT}=\frac{1}{\left|U_{t}\right|} \sum_{u \in U_{t}} \frac{|\{i, i \in(L(u) \cap \sim \Phi)\}|}{|L(u)|}\]

\(U_{t}\) is the number of users in the test set.

\(L_{u}\) is the recommended list of items for user u.

\(\sim \Phi\) medium-tail items.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [APLT]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of APLT

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.bias.arp package¶

Submodules¶

elliot.evaluation.metrics.bias.arp.arp module¶

This is the implementation of the Average Recommendation Popularity metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.bias.arp.arp.ARP(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Average Recommendation Popularity

This class represents the implementation of the Average Recommendation Popularity recommendation metric.

For further details, please refer to the paper

\[\mathrm {ARP}=\frac{1}{\left|U_{t}\right|} \sum_{u \in U_{t}} \frac{\sum_{i \in L_{u}} \phi(i)}{\left|L_{u}\right|}\]

\(U_{t}\) is the number of users in the test set.

\(L_{u}\) is the recommended list of items for user u.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [ARP]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of ARP

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.bias.pop_reo package¶

Submodules¶

elliot.evaluation.metrics.bias.pop_reo.extended_pop_reo module¶

This is the implementation of the Popularity-based Ranking-based Equal Opportunity (REO) metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.bias.pop_reo.extended_pop_reo.ExtendedPopREO(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended Popularity-based Ranking-based Equal Opportunity

This class represents the implementation of the Extended Popularity-based Ranking-based Equal Opportunity (REO) recommendation metric.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedPopREO

eval()[source]¶: Evaluation function :return: the overall averaged value of PopREO

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

elliot.evaluation.metrics.bias.pop_reo.pop_reo module¶

This is the implementation of the Popularity-based Ranking-based Equal Opportunity (REO) metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.bias.pop_reo.pop_reo.PopREO(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Popularity-based Ranking-based Equal Opportunity

This class represents the implementation of the Popularity-based Ranking-based Equal Opportunity (REO) recommendation metric.

For further details, please refer to the paper

\[\mathrm {REO}=\frac{{std}\left(P\left(R @ k \mid g=g_{1}, y=1\right) \ldots P\left(R(a) k=g_{A}, y=1\right)\right)} {{mean}\left(P\left(R @ k \mid g=g_{1}, y=1\right) \ldots P\left(R @ k \mid g=g_{A}, y=1\right)\right)}\]

\(P\left(R @ k \mid g=g_{a}, y=1\right) = \frac{\sum_{u=1}^{N} \sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right) Y\left(u, R_{u, i}\right)} {\sum_{u=1}^{N} \sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i) Y(u, i)}\)

\(Y\left(u, R_{u, i}\right)\) identifies the ground-truth label of a user-item pair left(u, R_{u, i}right), if item R_{u, i} is liked by user 𝑢, returns 1, otherwise 0

\(\sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right) Y\left(u, R_{u, i}\right)\) counts how many items in test set from group {g_a} are ranked in top-𝑘 for user u

\(\sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i) Y(u, i)\) counts the total number of items from group {g_a} 𝑎 in test set for user u

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [PopREO]

eval()[source]¶: Evaluation function :return: the overall averaged value of PopREO

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.bias.pop_rsp package¶

Submodules¶

elliot.evaluation.metrics.bias.pop_rsp.extended_pop_rsp module¶

This is the implementation of the Popularity-based Ranking-based Statistical Parity (RSP) metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.bias.pop_rsp.extended_pop_rsp.ExtendedPopRSP(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended Popularity-based Ranking-based Statistical Parity

This class represents the implementation of the Extended Popularity-based Ranking-based Statistical Parity (RSP) recommendation metric.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedPopRSP

eval()[source]¶: Evaluation function :return: the overall averaged value of PopRSP

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

elliot.evaluation.metrics.bias.pop_rsp.pop_rsp module¶

This is the implementation of the Popularity-based Ranking-based Statistical Parity (RSP) metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.bias.pop_rsp.pop_rsp.PopRSP(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Popularity-based Ranking-based Statistical Parity

This class represents the implementation of the Popularity-based Ranking-based Statistical Parity (RSP) recommendation metric.

For further details, please refer to the paper

\[\mathrm {RSP}=\frac{{std}\left(P\left(R @ k \mid g=g_{1}\right), \ldots, P\left(R @ k \mid g=g_{A}\right)\right)} {{mean}\left(P\left(R @ k \mid g=g_{1}\right), \ldots, P\left(R @ k \mid g=g_{A}\right)\right)}\]

:math P(R @ k mid g=g_{A})) = frac{sum_{u=1}^{N} sum_{i=1}^{k} G_{g_{a}}(R_{u, i})} {sum_{u=1}^{N} sum_{i in I backslash I_{u}^{+}} G_{g_{a}}(i)}

\(\sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right)\) calculates how many un-interacted items from group {g_a} are ranked in top-𝑘 for user u.

\(\sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i)\) calculates how many un-interacted items belong to group {g_a} for u

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [PopRSP]

eval()[source]¶: Evaluation function :return: the overall averaged value of PopRSP

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.coverage package¶

Subpackages¶

elliot.evaluation.metrics.coverage.item_coverage package¶

Submodules¶

elliot.evaluation.metrics.coverage.item_coverage.item_coverage module¶

This is the implementation of the Item Coverage metric. It directly proceeds from a system-wise computation, and it considers all the users at the same time.

class elliot.evaluation.metrics.coverage.item_coverage.item_coverage.ItemCoverage(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Item Coverage

This class represents the implementation of the Item Coverage recommendation metric.

For further details, please refer to the book

Note

The simplest measure of catalog coverage is the percentage of all items that can ever be recommended. This measure can be computed in many cases directly given the algorithm and the input data set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [ItemCoverage]

eval()[source]¶: Evaluation function :return: the overall averaged value of Item Coverage

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

This is the Item Coverage metric module.

This module contains and expose the recommendation metric.

elliot.evaluation.metrics.coverage.num_retrieved package¶

Submodules¶

elliot.evaluation.metrics.coverage.num_retrieved.num_retrieved module¶

This is the implementation of the NumRetrieved metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.coverage.num_retrieved.num_retrieved.NumRetrieved(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Number of Recommendations Retrieved

This class represents the implementation of the Number of Recommendations Retrieved recommendation metric.

For further details, please refer to the link

simple_metrics: [NumRetrieved]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of NumRetrieved

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.coverage.user_coverage package¶

Submodules¶

elliot.evaluation.metrics.coverage.user_coverage.user_coverage module¶

This is the implementation of the User Coverage metric. It directly proceeds from a system-wise computation, and it considers all the users at the same time.

class elliot.evaluation.metrics.coverage.user_coverage.user_coverage.UserCoverage(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

User Coverage

This class represents the implementation of the User Coverage recommendation metric.

For further details, please refer to the book

Note

The proportion of users or user interactions for which the system can recommend items. In many applications the recommender may not provide recommendations for some users due to, e.g. low confidence in the accuracy of predictions for that user.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [UserCoverage]

eval()[source]¶: Evaluation function :return: the overall averaged value of User Coverage

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

elliot.evaluation.metrics.coverage.user_coverage.user_coverage_at_n module¶

This is the implementation of the User Coverage metric. It directly proceeds from a system-wise computation, and it considers all the users at the same time.

class elliot.evaluation.metrics.coverage.user_coverage.user_coverage_at_n.UserCoverageAtN(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

User Coverage on Top-N rec. Lists

This class represents the implementation of the User Coverage recommendation metric.

For further details, please refer to the book

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [UserCoverageAtN]

eval()[source]¶: Evaluation function :return: the overall averaged value of User Coverage

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.diversity package¶

Subpackages¶

elliot.evaluation.metrics.diversity.SRecall package¶

Submodules¶

elliot.evaluation.metrics.diversity.SRecall.srecall module¶

This is the implementation of the SRecall metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.diversity.SRecall.srecall.SRecall(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Subtopic Recall

This class represents the implementation of the Subtopic Recall (S-Recall) recommendation metric.

For further details, please refer to the paper

\[\mathrm {SRecall}=\frac{\left|\cup_{i=1}^{K} {subtopics}\left(d_{i}\right)\right|}{n_{A}}\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [SRecall]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of SRecall

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.diversity.gini_index package¶

Submodules¶

elliot.evaluation.metrics.diversity.gini_index.gini_index module¶

This is the implementation of the Gini Index metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.diversity.gini_index.gini_index.GiniIndex(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Gini Index

This class represents the implementation of the Gini Index recommendation metric.

For further details, please refer to the book

\[\mathrm {GiniIndex}=\frac{1}{n-1} \sum_{j=1}^{n}(2 j-n-1) p\left(i_{j}\right)\]

\(i_{j}\) is the list of items ordered according to increasing p(i)

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [Gini]

eval()[source]¶: Evaluation function :return: the overall averaged value of Gini Index

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.diversity.shannon_entropy package¶

Submodules¶

elliot.evaluation.metrics.diversity.shannon_entropy.shannon_entropy module¶

This is the implementation of the Shannon Entropy metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.diversity.shannon_entropy.shannon_entropy.ShannonEntropy(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Shannon Entropy

This class represents the implementation of the Shannon Entropy recommendation metric.

For further details, please refer to the book

\[\mathrm {ShannonEntropy}=-\sum_{i=1}^{n} p(i) \log p(i)\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [SEntropy]

eval()[source]¶: Evaluation function :return: the overall value of Shannon Entropy

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.fairness package¶

Subpackages¶

elliot.evaluation.metrics.fairness.BiasDisparity package¶

Submodules¶

elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBD module¶

This is the implementation of the Bias Disparity metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBD.BiasDisparityBD(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Bias Disparity - Standard

This class represents the implementation of the Bias Disparity recommendation metric.

For further details, please refer to the paper

\[\mathrm {BD(G, C)}=\frac{B_{R}(G, C)-B_{S}(G, C)}{B_{S}(G, C)}\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
    - metric: BiasDisparityBD
      user_clustering_name: Happiness
      user_clustering_file: ../data/movielens_1m/u_happy.tsv
      item_clustering_name: ItemPopularity
      item_clustering_file: ../data/movielens_1m/i_pop.tsv

eval()[source]¶

get()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

process()[source]¶: Evaluation function :return: the overall value of Bias Disparity

elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBR module¶

This is the implementation of the Bias Disparity - Bias Recommendations metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBR.BiasDisparityBR(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Bias Disparity - Bias Recommendations

This class represents the implementation of the Bias Disparity - Bias Recommendations recommendation metric.

For further details, please refer to the paper

\[\mathrm {BD(G, C)}=\frac{B_{R}(G, C)-B_{S}(G, C)}{B_{S}(G, C)}\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
    - metric: BiasDisparityBR
      user_clustering_name: Happiness
      user_clustering_file: ../data/movielens_1m/u_happy.tsv
      item_clustering_name: ItemPopularity
      item_clustering_file: ../data/movielens_1m/i_pop.tsv

eval()[source]¶

get()[source]¶

get_BR()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

process()[source]¶: Evaluation function :return: the overall value of Bias Disparity - Bias Recommendations

elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBS module¶

This is the implementation of the Bias Disparity - Bias Source metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.BiasDisparity.BiasDisparityBS.BiasDisparityBS(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Bias Disparity - Bias Source

This class represents the implementation of the Bias Disparity - Bias Source recommendation metric.

For further details, please refer to the paper

\[\mathrm {B_{S}(G, C)}=\frac{P R_{S}(G, C)}{P(C)}\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
    - metric: BiasDisparityBS
      user_clustering_name: Happiness
      user_clustering_file: ../data/movielens_1m/u_happy.tsv
      item_clustering_name: ItemPopularity
      item_clustering_file: ../data/movielens_1m/i_pop.tsv

eval()[source]¶

get()[source]¶

get_BS()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

process()[source]¶: Evaluation function :return: the overall value of Bias Disparity - Bias Source

Module contents¶

elliot.evaluation.metrics.fairness.MAD package¶

Submodules¶

elliot.evaluation.metrics.fairness.MAD.ItemMADranking module¶

This is the implementation of the Item MAD ranking metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.MAD.ItemMADranking.ItemMADranking(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Item MAD Ranking-based

This class represents the implementation of the Item MAD ranking recommendation metric.

For further details, please refer to the paper

\[\mathrm {MAD}={avg}_{i, j}({MAD}(R^{(i)}, R^{(j)}))\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ItemMADranking
  clustering_name: ItemPopularity
  clustering_file: ../data/movielens_1m/i_pop.tsv

eval()[source]¶: Evaluation function :return: the overall averaged value of Item MAD ranking

get()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

elliot.evaluation.metrics.fairness.MAD.ItemMADrating module¶

This is the implementation of the Item MAD rating metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.MAD.ItemMADrating.ItemMADrating(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Item MAD Rating-based

This class represents the implementation of the Item MAD rating recommendation metric.

For further details, please refer to the paper

\[\mathrm {MAD}={avg}_{i, j}({MAD}(R^{(i)}, R^{(j)}))\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ItemMADrating
  clustering_name: ItemPopularity
  clustering_file: ../data/movielens_1m/i_pop.tsv

eval()[source]¶: Evaluation function :return: the overall averaged value of Item MAD rating

get()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

elliot.evaluation.metrics.fairness.MAD.UserMADranking module¶

This is the implementation of the User MAD ranking metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.MAD.UserMADranking.UserMADranking(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

User MAD Ranking-based

This class represents the implementation of the User MAD ranking recommendation metric.

For further details, please refer to the paper

\[\mathrm {MAD}={avg}_{i, j}({MAD}(R^{(i)}, R^{(j)}))\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: UserMADranking
  clustering_name: Happiness
  clustering_file: ../data/movielens_1m/u_happy.tsv

compute_idcg(user: int, cutoff: int) → float[source]¶: Method to compute Ideal Discounted Cumulative Gain :param gain_map: :param cutoff: :return:

compute_user_ndcg(user_recommendations: List, user: int, cutoff: int) → float[source]¶: Method to compute normalized Discounted Cumulative Gain :param sorted_item_predictions: :param gain_map: :param cutoff: :return:

eval()[source]¶: Evaluation function :return: the overall averaged value of User MAD ranking

get()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

elliot.evaluation.metrics.fairness.MAD.UserMADrating module¶

This is the implementation of the User MAD rating metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.MAD.UserMADrating.UserMADrating(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

User MAD Rating-based

This class represents the implementation of the User MAD rating recommendation metric.

For further details, please refer to the paper

\[\mathrm {MAD}={avg}_{i, j}({MAD}(R^{(i)}, R^{(j)}))\]

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: UserMADrating
  clustering_name: Happiness
  clustering_file: ../data/movielens_1m/u_happy.tsv

eval()[source]¶: Evaluation function :return: the overall averaged value of User MAD rating

get()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

This is the Precision metric module.

This module contains and expose the recommendation metric.

elliot.evaluation.metrics.fairness.reo package¶

Submodules¶

elliot.evaluation.metrics.fairness.reo.reo module¶

This is the implementation of the Ranking-based Equal Opportunity (REO) metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.reo.reo.REO(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Ranking-based Equal Opportunity

This class represents the implementation of the Ranking-based Equal Opportunity (REO) recommendation metric.

For further details, please refer to the paper

\[\mathrm {REO}=\frac{{std}\left(P\left(R @ k \mid g=g_{1}, y=1\right) \ldots P\left(R(a) k=g_{A}, y=1\right)\right)} {{mean}\left(P\left(R @ k \mid g=g_{1}, y=1\right) \ldots P\left(R @ k \mid g=g_{A}, y=1\right)\right)}\]

\(P\left(R @ k \mid g=g_{a}, y=1\right) = \frac{\sum_{u=1}^{N} \sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right) Y\left(u, R_{u, i}\right)} {\sum_{u=1}^{N} \sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i) Y(u, i)}\)

\(Y\left(u, R_{u, i}\right)\) identifies the ground-truth label of a user-item pair left(u, R_{u, i}right), if item R_{u, i} is liked by user 𝑢, returns 1, otherwise 0

\(\sum_{i=1}^{k} G_{g_{a}}\left(R_{u, i}\right) Y\left(u, R_{u, i}\right)\) counts how many items in test set from group {g_a} are ranked in top-𝑘 for user u

\(\sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i) Y(u, i)\) counts the total number of items from group {g_a} 𝑎 in test set for user u

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
 - metric: REO
  clustering_name: ItemPopularity
  clustering_file: ../data/movielens_1m/i_pop.tsv

eval()[source]¶

get()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

process()[source]¶: Evaluation function :return: the overall value of Ranking-based Equal Opportunity (REO)

Module contents¶

elliot.evaluation.metrics.fairness.rsp package¶

Submodules¶

elliot.evaluation.metrics.fairness.rsp.rsp module¶

This is the implementation of the Ranking-based Statistical Parity (RSP) metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.fairness.rsp.rsp.RSP(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Ranking-based Statistical Parity

This class represents the implementation of the Ranking-based Statistical Parity (RSP) recommendation metric.

For further details, please refer to the paper

\[\mathrm {RSP}=\frac{{std}(P(R @ k \mid g=g_{1}), \ldots, P(R @ k \mid g=g_{A}))} {{mean}(P(R @ k \mid g=g_{1}), \ldots, P(R @ k \mid g=g_{A}))}\]

\(P(R @ k \mid g=g_{A})) = \frac{\sum_{u=1}^{N} \sum_{i=1}^{k} G_{g_{a}}(R_{u, i})} {\sum_{u=1}^{N} \sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i)}\)

\(\sum_{i=1}^{k} G_{g_{a}}(R_{u, i})\) calculates how many un-interacted items from group {g_a} are ranked in top-𝑘 for user u.

\(\sum_{i \in I \backslash I_{u}^{+}} G_{g_{a}}(i)\) calculates how many un-interacted items belong to group {g_a} for u

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
 - metric: RSP
  clustering_name: ItemPopularity
  clustering_file: ../data/movielens_1m/i_pop.tsv

eval()[source]¶

get()[source]¶

name()[source]¶: Metric Name Getter :return: returns the public name of the metric

process()[source]¶: Evaluation function :return: the overall value of Ranking-based Statistical Parity (RSP)

Module contents¶

elliot.evaluation.metrics.novelty package¶

Subpackages¶

elliot.evaluation.metrics.novelty.EFD package¶

Submodules¶

elliot.evaluation.metrics.novelty.EFD.efd module¶

This is the implementation of the Expected Free Discovery metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.novelty.EFD.efd.EFD(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Expected Free Discovery (EFD)

This class represents the implementation of the Expected Free Discovery recommendation metric.

For further details, please refer to the paper

Note

EFD can be read as the expected ICF of seen recommended items

\[\mathrm {EFD}=C \sum_{i_{k} \in R} {disc}(k) p({rel} \mid i_{k}, u)( -\log _{2} p(i \mid {seen}, \theta))\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [EFD]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Expected Free Discovery per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

elliot.evaluation.metrics.novelty.EFD.extended_efd module¶

This is the implementation of the Expected Free Discovery metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.novelty.EFD.extended_efd.ExtendedEFD(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended EFD

This class represents the implementation of the Extended Expected Free Discovery recommendation metric.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedEFD

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Expected Free Discovery per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.novelty.EPC package¶

Submodules¶

elliot.evaluation.metrics.novelty.EPC.epc module¶

This is the implementation of the Expected Popularity Complement metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.novelty.EPC.epc.EPC(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Expected Popularity Complement (EPC)

This class represents the implementation of the Expected Popularity Complement recommendation metric.

For further details, please refer to the paper

Note

EPC can be read as the expected number of seen relevant recommended items not previously seen

\[\mathrm{EPC}=C \sum_{i_{k} \in R} \operatorname{disc}(k) p\left(r e l \mid i_{k}, u\right)\left(1-p\left(\operatorname{seen} \mid t_{k}\right)\right)\]

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [EPC]

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Expected Popularity Complement per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

elliot.evaluation.metrics.novelty.EPC.extended_epc module¶

This is the implementation of the Expected Popularity Complement metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.novelty.EPC.extended_epc.ExtendedEPC(recommendations, config, params, eval_objects, additional_data)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Extended EPC

This class represents the implementation of the Extended EPC recommendation metric.

For further details, please refer to the paper

To compute the metric, add it to the config file adopting the following pattern:

complex_metrics:
- metric: ExtendedEPC

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Expected Popularity Complement per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

Module contents¶

elliot.evaluation.metrics.rating package¶

Subpackages¶

elliot.evaluation.metrics.rating.mae package¶

Submodules¶

elliot.evaluation.metrics.rating.mae.mae module¶

This is the implementation of the Mean Absolute Error metric. It proceeds from a system-wise computation.

class elliot.evaluation.metrics.rating.mae.mae.MAE(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Absolute Error

This class represents the implementation of the Mean Absolute Error recommendation metric.

For further details, please refer to the link

\[\mathrm{MAE}=\frac{1}{|{T}|} \sum_{(u, i) \in {T}}\left|\hat{r}_{u i}-r_{u i}\right|\]

\(T\) is the test set, \(\hat{r}_{u i}\) is the score predicted by the model,

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MAE]

eval()[source]¶: Evaluation function :return: the overall averaged value of Mean Absolute Error

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Mean Absolute Error per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

static needs_full_recommendations()[source]¶

Module contents¶

elliot.evaluation.metrics.rating.mse package¶

Submodules¶

elliot.evaluation.metrics.rating.mse.mse module¶

This is the implementation of the Mean Squared Error metric. It proceeds from a system-wise computation.

class elliot.evaluation.metrics.rating.mse.mse.MSE(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Mean Squared Error

This class represents the implementation of the Mean Squared Error recommendation metric.

For further details, please refer to the link

\[\mathrm{MSE} = \frac{1}{|{T}|} \sum_{(u, i) \in {T}}(\hat{r}_{u i}-r_{u i})^{2}\]

\(T\) is the test set, \(\hat{r}_{u i}\) is the score predicted by the model

\(r_{u i}\) the actual score of the test set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [MSE]

eval()[source]¶: Evaluation function :return: the overall averaged value of Mean Squared Error

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Mean Squared Error per user

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

static needs_full_recommendations()[source]¶

Module contents¶

elliot.evaluation.metrics.rating.rmse package¶

Submodules¶

elliot.evaluation.metrics.rating.rmse.rmse module¶

This is the implementation of the Root Mean Squared Error metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.rating.rmse.rmse.RMSE(recommendations, config, params, eval_objects)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

Root Mean Squared Error

This class represents the implementation of the Root Mean Squared Error recommendation metric.

For further details, please refer to the link

\[\mathrm{RMSE} = \sqrt{\frac{1}{|{T}|} \sum_{(u, i) \in {T}}(\hat{r}_{u i}-r_{u i})^{2}}\]

\(T\) is the test set, \(\hat{r}_{u i}\) is the score predicted by the model

\(r_{u i}\) the actual score of the test set.

To compute the metric, add it to the config file adopting the following pattern:

simple_metrics: [RMSE]

eval()[source]¶: Evaluation function :return: the overall averaged value of Root Mean Squared Error

eval_user_metric()[source]¶: Evaluation function :return: the overall averaged value of Root Mean Squared Error

static name()[source]¶: Metric Name Getter :return: returns the public name of the metric

static needs_full_recommendations()[source]¶

Module contents¶

Submodules¶

elliot.evaluation.metrics.base_metric module¶

This is the implementation of the Precision metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.base_metric.BaseMetric(recommendations, config, params, evaluation_objects, additional_data=None)[source]¶

Bases: abc.ABC

This class represents the implementation of the Precision recommendation metric. Passing ‘Precision’ to the metrics list will enable the computation of the metric.

eval()[source]¶

get()[source]¶

abstract name()[source]¶

static needs_full_recommendations()[source]¶

elliot.evaluation.metrics.metrics_utils module¶

class elliot.evaluation.metrics.metrics_utils.ProxyMetric(name='ProxyMetric', val=0, needs_full_recommendations=False)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

eval()[source]¶

name()[source]¶

needs_full_recommendations()[source]¶

class elliot.evaluation.metrics.metrics_utils.ProxyStatisticalMetric(name='ProxyMetric', val=0, user_val=0, needs_full_recommendations=False)[source]¶

Bases: elliot.evaluation.metrics.base_metric.BaseMetric

eval()[source]¶

eval_user_metric()[source]¶

name()[source]¶

needs_full_recommendations()[source]¶

elliot.evaluation.metrics.statistical_array_metric module¶

This is the implementation of the Precision metric. It proceeds from a user-wise computation, and average the values over the users.

class elliot.evaluation.metrics.statistical_array_metric.StatisticalMetric[source]¶

Bases: object

This class represents the implementation of the Precision recommendation metric. Passing ‘Precision’ to the metrics list will enable the computation of the metric.

abstract eval_user_metric()[source]¶

Module contents¶

This is the metrics’ module.

This module contains and expose the recommendation metrics. Each metric is encapsulated in a specific package.

See the implementation of Precision metric for creating new per-user metrics. See the implementation of Item Coverage for creating new cross-user metrics.

elliot.evaluation.metrics.parse_metric(metric)[source]¶

elliot.evaluation.metrics.parse_metrics(metrics)[source]¶

elliot.evaluation.popularity_utils package¶

Submodules¶

elliot.evaluation.popularity_utils.popularity module¶

Module description: This module provides a popularity class based on number of users who have experienced an item (user-item repetitions in the dataset are counted once)

class elliot.evaluation.popularity_utils.popularity.Popularity(data, pop_ratio=0.8)[source]¶

Bases: object

get_custom_pop_obj(pop_ratio=0.8)[source]¶

get_long_tail()[source]¶

get_pop_items()[source]¶

get_short_head()[source]¶

get_sorted_pop_items()[source]¶

Module contents¶

elliot.evaluation.relevance package¶

Submodules¶

elliot.evaluation.relevance.relevance module¶

Module description:

class elliot.evaluation.relevance.relevance.AbstractRelevanceSingleton[source]¶

Bases: abc.ABC

abstract get_rel(user, item)[source]¶

static logarithmic_ranking_discount(k: int) → float[source]¶: Method to compute logarithmic discount :param k: :return:

class elliot.evaluation.relevance.relevance.BinaryRelevance(test, rel_threshold)[source]¶

Bases: elliot.evaluation.relevance.relevance.AbstractRelevanceSingleton

get_rel(user, item)[source]¶

get_user_rel(user)[source]¶

get_user_rel_gains(user)[source]¶

class elliot.evaluation.relevance.relevance.DiscountedRelevance(test, rel_threshold)[source]¶

Bases: elliot.evaluation.relevance.relevance.AbstractRelevanceSingleton

get_rel(user, item)[source]¶

get_user_rel(user)[source]¶

get_user_rel_gains(user)[source]¶

class elliot.evaluation.relevance.relevance.Relevance(test, rel_threshold)[source]¶

Bases: object

property binary_relevance¶

property discounted_relevance¶

get_test()[source]¶

Module contents¶

Module description:

Submodules¶

elliot.evaluation.evaluator module¶

Module description:

class elliot.evaluation.evaluator.Evaluator(data: elliot.dataset.dataset.DataSet, params: types.SimpleNamespace)[source]¶

Bases: object

eval(recommendations)[source]¶: Runtime Evaluation of Accuracy Performance (top-k) :return:

eval_at_k(recommendations, k)[source]¶

get_needed_recommendations()[source]¶

elliot.evaluation.statistical_significance module¶

Module description:

class elliot.evaluation.statistical_significance.PairedTTest[source]¶

Bases: object

static common_users(arr_0: Dict[int, float], arr_1: Dict[int, float])[source]¶

static compare(arr_0: Dict[int, float], arr_1: Dict[int, float], users: List[int])[source]¶

class elliot.evaluation.statistical_significance.WilcoxonTest[source]¶

Bases: object

static common_users(arr_0: Dict[int, float], arr_1: Dict[int, float])[source]¶

static compare(arr_0: Dict[int, float], arr_1: Dict[int, float], users: List[int])[source]¶

Module contents¶

Module description:

elliot.hyperoptimization package¶

Submodules¶

elliot.hyperoptimization.model_coordinator module¶

Module description:

class elliot.hyperoptimization.model_coordinator.ModelCoordinator(data_objs, base: types.SimpleNamespace, params, model_class: ClassVar)[source]¶

Bases: object

This class handles the selection of hyperparameters for the hyperparameter tuning realized with HyperOpt.

objective(args)[source]¶: This function respect the signature, and the return format required for HyperOpt optimization :param args: a Dictionary that contains the new hyper-parameter values that will be used in the current run :return: it returns a Dictionary with loss, and status being required by HyperOpt, and params, and results being required by the framework

single()[source]¶: This function respect the signature, and the return format required for HyperOpt optimization :param args: a Dictionary that contains the new hyper-parameter values that will be used in the current run :return: it returns a Dictionary with loss, and status being required by HyperOpt, and params, and results being required by the framework

Module contents¶

Module description:

elliot.hyperoptimization.parse_algorithms(opt_alg)[source]¶

elliot.hyperoptimization.suggest(new_ids, domain, trials, seed, nbMaxSucessiveFailures=1000)[source]¶

elliot.namespace package¶

Submodules¶

elliot.namespace.namespace_model module¶

Module description:

class elliot.namespace.namespace_model.NameSpaceModel(config_path, base_folder_path_elliot, base_folder_path_config)[source]¶

Bases: object

fill_base()[source]¶

fill_model()[source]¶

elliot.namespace.namespace_model_builder module¶

Module description:

class elliot.namespace.namespace_model_builder.Builder[source]¶

Bases: abc.ABC

The Builder interface specifies methods for creating the different parts of the Product objects.

abstract property base¶

abstract models() → None[source]¶

class elliot.namespace.namespace_model_builder.NameSpaceBuilder(config_path, base_folder_path_elliot, base_folder_path_config)[source]¶

Bases: elliot.namespace.namespace_model_builder.Builder

property base¶

models() → tuple[source]¶

Module contents¶

Module description:

elliot.prefiltering package¶

Submodules¶

elliot.prefiltering.standard_prefilters module¶

class elliot.prefiltering.standard_prefilters.PreFilter[source]¶

Bases: object

static filter(d: pandas.DataFrame, ns: types.SimpleNamespace) → pandas.DataFrame[source]¶

static filter_items_by_popularity(d: pandas.DataFrame, threshold) → pandas.DataFrame[source]¶

static filter_iterative_k_core(d: pandas.DataFrame, threshold) → pandas.DataFrame[source]¶

static filter_ratings_by_global_average(d: pandas.DataFrame) → pandas.DataFrame[source]¶

static filter_ratings_by_threshold(d: pandas.DataFrame, threshold) → pandas.DataFrame[source]¶

static filter_ratings_by_user_average(d: pandas.DataFrame) → pandas.DataFrame[source]¶

static filter_retain_cold_users(d: pandas.DataFrame, threshold) → pandas.DataFrame[source]¶

static filter_rounds_k_core(d: pandas.DataFrame, threshold, n_rounds) → pandas.DataFrame[source]¶

static filter_users_by_profile_size(d: pandas.DataFrame, threshold) → pandas.DataFrame[source]¶

Module contents¶

Module description:

elliot.recommender package¶

Subpackages¶

elliot.recommender.NN package¶

Subpackages¶

elliot.recommender.NN.attribute_item_knn package¶

Submodules¶

elliot.recommender.NN.attribute_item_knn.attribute_item_knn module¶

Module description:

class elliot.recommender.NN.attribute_item_knn.attribute_item_knn.AttributeItemKNN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Attribute Item-kNN proposed in MyMediaLite Recommender System Library

For further details, please refer to the paper

Parameters

neighbors – Number of item neighbors
similarity – Similarity function

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AttributeItemKNN:
    meta:
      save_recs: True
    neighbors: 40
    similarity: cosine

build_feature_sparse()[source]¶

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.NN.attribute_item_knn.attribute_item_knn_similarity module¶

class elliot.recommender.NN.attribute_item_knn.attribute_item_knn_similarity.Similarity(data, attribute_matrix, num_neighbors, similarity)[source]¶

Bases: object

Simple kNN class

compute_cosine(i_index, j_index)[source]¶

compute_neighbors()[source]¶

get_item_neighbors(item)[source]¶

get_model_state()[source]¶

get_transactions()[source]¶

get_user_recs(u, k)[source]¶

initialize()[source]¶: This function initialize the data model

process_similarity(similarity)[source]¶

static score_item(neighs, user_items)[source]¶

set_model_state(saving_dict)[source]¶

Module contents¶

elliot.recommender.NN.attribute_user_knn package¶

Submodules¶

elliot.recommender.NN.attribute_user_knn.attribute_user_knn module¶

Module description:

class elliot.recommender.NN.attribute_user_knn.attribute_user_knn.AttributeUserKNN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Attribute User-kNN proposed in MyMediaLite Recommender System Library

For further details, please refer to the paper

Parameters

neighbors – Number of item neighbors
similarity – Similarity function
profile – Profile type (‘binary’, ‘tfidf’)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AttributeUserKNN:
    meta:
      save_recs: True
    neighbors: 40
    similarity: cosine
    profile: binary

build_feature_sparse()[source]¶

build_feature_sparse_values()[source]¶

compute_binary_profile(user_items_dict: Dict)[source]¶

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.NN.attribute_user_knn.attribute_user_knn_similarity module¶

class elliot.recommender.NN.attribute_user_knn.attribute_user_knn_similarity.Similarity(data, attribute_matrix, num_neighbors, similarity)[source]¶

Bases: object

Simple kNN class

compute_neighbors()[source]¶

get_model_state()[source]¶

get_transactions()[source]¶

get_user_neighbors(item)[source]¶

get_user_recs(u, k)[source]¶

initialize()[source]¶: This function initialize the data model

process_similarity(similarity)[source]¶

static score_item(neighs, user_neighs_items)[source]¶

set_model_state(saving_dict)[source]¶

elliot.recommender.NN.attribute_user_knn.tfidf_utils module¶

class elliot.recommender.NN.attribute_user_knn.tfidf_utils.TFIDF(map: Dict[int, List[int]])[source]¶

Bases: object

get_profiles(ratings: Dict[int, Dict[int, float]])[source]¶

tfidf()[source]¶

Module contents¶

elliot.recommender.NN.item_knn package¶

Submodules¶

elliot.recommender.NN.item_knn.aiolli_ferrari module¶

Created on 23/10/17 @author: Maurizio Ferrari Dacrema

class elliot.recommender.NN.item_knn.aiolli_ferrari.AiolliSimilarity(data, maxk=40, shrink=100, similarity='cosine', normalize=True)[source]¶

Bases: object

get_user_recs(user, k=100)[source]¶

initialize()[source]¶

predict(u, i)[source]¶

class elliot.recommender.NN.item_knn.aiolli_ferrari.Compute_Similarity(dataMatrix, topK=100, shrink=0, normalize=True, asymmetric_alpha=0.5, tversky_alpha=1.0, tversky_beta=1.0, similarity='cosine', row_weights=None)[source]¶

Bases: object

applyAdjustedCosine()[source]¶: Remove from every data point the average for the corresponding row :return:

applyPearsonCorrelation()[source]¶: Remove from every data point the average for the corresponding column :return:

compute_similarity(start_col=None, end_col=None, block_size=100)[source]¶: Compute the similarity for the given dataset :param self: :param start_col: column to begin with :param end_col: column to stop before, end_col is excluded :return:

useOnlyBooleanInteractions()[source]¶

elliot.recommender.NN.item_knn.aiolli_ferrari.check_matrix(X, format='csc', dtype=<class 'numpy.float32'>)[source]¶: This function takes a matrix as input and transforms it into the specified format. The matrix in input can be either sparse or ndarray. If the matrix in input has already the desired format, it is returned as-is the dtype parameter is always applied and the default is np.float32 :param X: :param format: :param dtype: :return:

elliot.recommender.NN.item_knn.item_knn module¶

Module description:

class elliot.recommender.NN.item_knn.item_knn.ItemKNN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Amazon.com recommendations: item-to-item collaborative filtering

For further details, please refer to the paper

Parameters

neighbors – Number of item neighbors
similarity – Similarity function
implementation – Implementation type (‘aiolli’, ‘classical’)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ItemKNN:
    meta:
      save_recs: True
    neighbors: 40
    similarity: cosine
    implementation: aiolli

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.NN.item_knn.item_knn_similarity module¶

class elliot.recommender.NN.item_knn.item_knn_similarity.Similarity(data, num_neighbors, similarity)[source]¶

Bases: object

Simple kNN class

compute_cosine(i_index, j_index)[source]¶

compute_neighbors()[source]¶

get_item_neighbors(item)[source]¶

get_model_state()[source]¶

get_transactions()[source]¶

get_user_recs(u, k)[source]¶

initialize()[source]¶: This function initialize the data model

process_cosine()[source]¶

process_similarity(similarity)[source]¶

static score_item(neighs, user_items)[source]¶

set_model_state(saving_dict)[source]¶

Module contents¶

elliot.recommender.NN.user_knn package¶

Submodules¶

elliot.recommender.NN.user_knn.aiolli_ferrari module¶

Created on 23/10/17 @author: Maurizio Ferrari Dacrema

class elliot.recommender.NN.user_knn.aiolli_ferrari.AiolliSimilarity(data, maxk=40, shrink=100, similarity='cosine', normalize=True)[source]¶

Bases: object

get_user_recs(user, k=100)[source]¶

initialize()[source]¶

predict(u, i)[source]¶

class elliot.recommender.NN.user_knn.aiolli_ferrari.Compute_Similarity(dataMatrix, topK=100, shrink=0, normalize=True, asymmetric_alpha=0.5, tversky_alpha=1.0, tversky_beta=1.0, similarity='cosine', row_weights=None)[source]¶

Bases: object

applyAdjustedCosine()[source]¶: Remove from every data point the average for the corresponding row :return:

applyPearsonCorrelation()[source]¶: Remove from every data point the average for the corresponding column :return:

compute_similarity(start_col=None, end_col=None, block_size=100)[source]¶: Compute the similarity for the given dataset :param self: :param start_col: column to begin with :param end_col: column to stop before, end_col is excluded :return:

useOnlyBooleanInteractions()[source]¶

elliot.recommender.NN.user_knn.aiolli_ferrari.check_matrix(X, format='csc', dtype=<class 'numpy.float32'>)[source]¶: This function takes a matrix as input and transforms it into the specified format. The matrix in input can be either sparse or ndarray. If the matrix in input has already the desired format, it is returned as-is the dtype parameter is always applied and the default is np.float32 :param X: :param format: :param dtype: :return:

elliot.recommender.NN.user_knn.user_knn module¶

Module description:

class elliot.recommender.NN.user_knn.user_knn.UserKNN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

GroupLens: An Open Architecture for Collaborative Filtering of Netnews

For further details, please refer to the paper

Parameters

neighbors – Number of item neighbors
similarity – Similarity function
implementation – Implementation type (‘aiolli’, ‘classical’)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  UserKNN:
    meta:
      save_recs: True
    neighbors: 40
    similarity: cosine
    implementation: aiolli

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.NN.user_knn.user_knn_similarity module¶

class elliot.recommender.NN.user_knn.user_knn_similarity.Similarity(data, num_neighbors, similarity)[source]¶

Bases: object

Simple kNN class

compute_neighbors()[source]¶

get_model_state()[source]¶

get_transactions()[source]¶

get_user_neighbors(item)[source]¶

get_user_recs(u, k)[source]¶

initialize()[source]¶: This function initialize the data model

process_similarity(similarity)[source]¶

static score_item(neighs, user_neighs_items)[source]¶

set_model_state(saving_dict)[source]¶

Module contents¶

elliot.recommender.adversarial package¶

Subpackages¶

elliot.recommender.adversarial.AMF package¶

Submodules¶

elliot.recommender.adversarial.AMF.AMF module¶

Module description:

class elliot.recommender.adversarial.AMF.AMF.AMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Adversarial Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factor
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
eps – Perturbation Budget
l_adv – Adversarial regularization coefficient
adversarial_epochs – Adversarial epochs

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 200
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    eps: 0.1
    l_adv: 0.001
    adversarial_epochs: 10

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.adversarial.AMF.AMF_model module¶

Module description:

class elliot.recommender.adversarial.AMF.AMF_model.AMF_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

build_perturbation(batch)[source]¶: Evaluate Adversarial Perturbation with FGSM-like Approach

call(inputs, training=None)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_positions(predictions, train_mask, items, inner_test_user_true_mask)[source]¶

get_top_k(predictions, train_mask, k=100)[source]¶

predict(start, stop, **kwargs)[source]¶

Generates output predictions for the input samples.

Computation is done in batches. This method is designed for performance in large scale inputs. For small amount of inputs that fit in one batch, directly using __call__ is recommended for faster execution, e.g., model(x), or model(x, training=False) if you have layers such as tf.keras.layers.BatchNormalization that behaves differently during inference. Also, note the fact that test loss is not affected by regularization layers like noise and dropout.

Parameters

x –
Input samples. It could be: - A Numpy array (or array-like), or a list of arrays

(in case the model has multiple inputs).
- A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).
- A tf.data dataset.
- A generator or keras.utils.Sequence instance.
A more detailed description of unpacking behavior for iterator types (Dataset, generator, Sequence) is given in the Unpacking behavior for iterator-like inputs section of Model.fit.
batch_size – Integer or None. Number of samples per batch. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of dataset, generators, or keras.utils.Sequence instances (since they generate batches).
verbose – Verbosity mode, 0 or 1.
steps – Total number of steps (batches of samples) before declaring the prediction round finished. Ignored with the default value of None. If x is a tf.data dataset and steps is None, predict will run until the input dataset is exhausted.
callbacks – List of keras.callbacks.Callback instances. List of callbacks to apply during prediction. See [callbacks](/api_docs/python/tf/keras/callbacks).
max_queue_size – Integer. Used for generator or keras.utils.Sequence input only. Maximum size for the generator queue. If unspecified, max_queue_size will default to 10.
workers – Integer. Used for generator or keras.utils.Sequence input only. Maximum number of processes to spin up when using process-based threading. If unspecified, workers will default to 1. If 0, will execute the generator on the main thread.
use_multiprocessing – Boolean. Used for generator or keras.utils.Sequence input only. If True, use process-based threading. If unspecified, use_multiprocessing will default to False. Note that because this implementation relies on multiprocessing, you should not pass non-picklable arguments to the generator as they can’t be passed easily to children processes.

See the discussion of Unpacking behavior for iterator-like inputs for Model.fit. Note that Model.predict uses the same interpretation rules as Model.fit and Model.evaluate, so inputs must be unambiguous for all three methods.

Returns

Numpy array(s) of predictions.

Raises

RuntimeError – If model.predict is wrapped in tf.function.
ValueError – In case of mismatch between the provided input data and the model’s expectations, or in case a stateful model receives a number of samples that is not a multiple of the batch size.

train_step(batch, user_adv_train=False)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

Module contents¶

Module description:

elliot.recommender.adversarial.AMR package¶

Submodules¶

elliot.recommender.adversarial.AMR.AMR module¶

Module description:

class elliot.recommender.adversarial.AMR.AMR.AMR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Adversarial Multimedia Recommender

For further details, please refer to the paper

Parameters

factors – Number of latent factor
factors_d – Image-feature dimensionality
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_e – Regularization coefficient of image matrix embedding
eps – Perturbation Budget
l_adv – Adversarial regularization coefficient
adversarial_epochs – Adversarial epochs

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AMR:
    meta:
      save_recs: True
    epochs: 10
    factors: 200
    factors_d: 20
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    l_e: 0.1
    eps: 0.1
    l_adv: 0.001
    adversarial_epochs: 5

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.adversarial.AMR.AMR_model module¶

Module description:

class elliot.recommender.adversarial.AMR.AMR_model.AMR_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

build_perturbation(batch)[source]¶: Evaluate Adversarial Perturbation with FGSM-like Approach

call(inputs, training=None, mask=None)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(start, stop)[source]¶

Generates output predictions for the input samples.

Computation is done in batches. This method is designed for performance in large scale inputs. For small amount of inputs that fit in one batch, directly using __call__ is recommended for faster execution, e.g., model(x), or model(x, training=False) if you have layers such as tf.keras.layers.BatchNormalization that behaves differently during inference. Also, note the fact that test loss is not affected by regularization layers like noise and dropout.

Parameters

x –
Input samples. It could be: - A Numpy array (or array-like), or a list of arrays

(in case the model has multiple inputs).
- A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).
- A tf.data dataset.
- A generator or keras.utils.Sequence instance.
A more detailed description of unpacking behavior for iterator types (Dataset, generator, Sequence) is given in the Unpacking behavior for iterator-like inputs section of Model.fit.
batch_size – Integer or None. Number of samples per batch. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of dataset, generators, or keras.utils.Sequence instances (since they generate batches).
verbose – Verbosity mode, 0 or 1.
steps – Total number of steps (batches of samples) before declaring the prediction round finished. Ignored with the default value of None. If x is a tf.data dataset and steps is None, predict will run until the input dataset is exhausted.
callbacks – List of keras.callbacks.Callback instances. List of callbacks to apply during prediction. See [callbacks](/api_docs/python/tf/keras/callbacks).
max_queue_size – Integer. Used for generator or keras.utils.Sequence input only. Maximum size for the generator queue. If unspecified, max_queue_size will default to 10.
workers – Integer. Used for generator or keras.utils.Sequence input only. Maximum number of processes to spin up when using process-based threading. If unspecified, workers will default to 1. If 0, will execute the generator on the main thread.
use_multiprocessing – Boolean. Used for generator or keras.utils.Sequence input only. If True, use process-based threading. If unspecified, use_multiprocessing will default to False. Note that because this implementation relies on multiprocessing, you should not pass non-picklable arguments to the generator as they can’t be passed easily to children processes.

See the discussion of Unpacking behavior for iterator-like inputs for Model.fit. Note that Model.predict uses the same interpretation rules as Model.fit and Model.evaluate, so inputs must be unambiguous for all three methods.

Returns

Numpy array(s) of predictions.

Raises

RuntimeError – If model.predict is wrapped in tf.function.
ValueError – In case of mismatch between the provided input data and the model’s expectations, or in case a stateful model receives a number of samples that is not a multiple of the batch size.

train_step(batch, user_adv_train=False)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

Module contents¶

Module description:

Module contents¶

elliot.recommender.algebric package¶

Subpackages¶

elliot.recommender.algebric.slope_one package¶

Submodules¶

elliot.recommender.algebric.slope_one.slope_one module¶

Module description: Lemire, Daniel, and Anna Maclachlan. “Slope one predictors for online rating-based collaborative filtering.” Proceedings of the 2005 SIAM International Conference on Data Mining. Society for Industrial and Applied Mathematics

class elliot.recommender.algebric.slope_one.slope_one.SlopeOne(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Slope One Predictors for Online Rating-Based Collaborative Filtering

For further details, please refer to the paper

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  SlopeOne:
    meta:
      save_recs: True

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.algebric.slope_one.slope_one_model module¶

Lemire, Daniel, and Anna Maclachlan. “Slope one predictors for online rating-based collaborative filtering.” Proceedings of the 2005 SIAM International Conference on Data Mining. Society for Industrial and Applied Mathematics

class elliot.recommender.algebric.slope_one.slope_one_model.SlopeOneModel(data)[source]¶

Bases: object

get_model_state()[source]¶

get_user_recs(u, k)[source]¶

initialize()[source]¶

predict(user, item)[source]¶

set_model_state(saving_dict)[source]¶

Module contents¶

elliot.recommender.autoencoders package¶

Subpackages¶

elliot.recommender.autoencoders.dae package¶

Submodules¶

elliot.recommender.autoencoders.dae.multi_dae module¶

Module description:

class elliot.recommender.autoencoders.dae.multi_dae.MultiDAE(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Variational Autoencoders for Collaborative Filtering

For further details, please refer to the paper

Parameters

intermediate_dim – Number of intermediate dimension
latent_dim – Number of latent factors
reg_lambda – Regularization coefficient
lr – Learning rate
dropout_pkeep – Dropout probaility

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  MultiDAE:
    meta:
      save_recs: True
    epochs: 10
    intermediate_dim: 600
    latent_dim: 200
    reg_lambda: 0.01
    lr: 0.001
    dropout_pkeep: 1

property name¶

train()[source]¶

elliot.recommender.autoencoders.dae.multi_dae_model module¶

Module description:

class elliot.recommender.autoencoders.dae.multi_dae_model.Decoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

Converts z, the encoded vector, back into a uaser interaction vector.

call(inputs, **kwargs)[source]¶

class elliot.recommender.autoencoders.dae.multi_dae_model.DenoisingAutoEncoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

Combines the encoder and decoder into an end-to-end model for training.

call(inputs, training=None, **kwargs)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

class elliot.recommender.autoencoders.dae.multi_dae_model.Encoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

Maps user-item interactions to a triplet (z_mean, z_log_var, z).

call(inputs, training=None)[source]¶

Module contents¶

Module description:

elliot.recommender.autoencoders.vae package¶

Submodules¶

elliot.recommender.autoencoders.vae.multi_vae module¶

Module description:

class elliot.recommender.autoencoders.vae.multi_vae.MultiVAE(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Variational Autoencoders for Collaborative Filtering

For further details, please refer to the paper

Parameters

intermediate_dim – Number of intermediate dimension
latent_dim – Number of latent factors
reg_lambda – Regularization coefficient
lr – Learning rate
dropout_pkeep – Dropout probaility

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  MultiVAE:
    meta:
      save_recs: True
    epochs: 10
    intermediate_dim: 600
    latent_dim: 200
    reg_lambda: 0.01
    lr: 0.001
    dropout_pkeep: 1

property name¶

train()[source]¶

elliot.recommender.autoencoders.vae.multi_vae_model module¶

Module description:

class elliot.recommender.autoencoders.vae.multi_vae_model.Decoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

Converts z, the encoded digit vector, back into a readable digit.

call(inputs, **kwargs)[source]¶

class elliot.recommender.autoencoders.vae.multi_vae_model.Encoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

Maps MNIST digits to a triplet (z_mean, z_log_var, z).

call(inputs, training=None)[source]¶

class elliot.recommender.autoencoders.vae.multi_vae_model.Sampling(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

Uses (z_mean, z_log_var) to sample z, the vector encoding a digit.

call(inputs)[source]¶

class elliot.recommender.autoencoders.vae.multi_vae_model.VariationalAutoEncoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

Combines the encoder and decoder into an end-to-end model for training.

call(inputs, training=None, **kwargs)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch, anneal_ph=0.0, **kwargs)[source]¶

Module contents¶

Module description:

Module contents¶

Module description:

elliot.recommender.content_based package¶

Subpackages¶

elliot.recommender.content_based.VSM package¶

Submodules¶

elliot.recommender.content_based.VSM.tfidf_utils module¶

class elliot.recommender.content_based.VSM.tfidf_utils.TFIDF(map: Dict[int, List[int]])[source]¶

Bases: object

get_profiles(ratings: Dict[int, Dict[int, float]])[source]¶

tfidf()[source]¶

elliot.recommender.content_based.VSM.vector_space_model module¶

Module description:

class elliot.recommender.content_based.VSM.vector_space_model.VSM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Vector Space Model

For further details, please refer to the paper and the paper

Parameters

similarity – Similarity metric
user_profile –
item_profile –

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  VSM:
    meta:
      save_recs: True
    similarity: cosine
    user_profile: binary
    item_profile: binary

build_feature_sparse(feature_dict, num_entities)[source]¶

build_feature_sparse_values(feature_dict, num_entities)[source]¶

compute_binary_profile(user_items_dict: Dict)[source]¶

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.content_based.VSM.vector_space_model_similarity module¶

class elliot.recommender.content_based.VSM.vector_space_model_similarity.Similarity(data, user_profile_matrix, item_attribute_matrix, similarity)[source]¶

Bases: object

Simple kNN class

get_model_state()[source]¶

get_transactions()[source]¶

get_user_recs(u, k)[source]¶

initialize()[source]¶: This function initialize the data model

process_similarity(similarity)[source]¶

set_model_state(saving_dict)[source]¶

Module contents¶

elliot.recommender.gan package¶

Subpackages¶

elliot.recommender.gan.CFGAN package¶

Submodules¶

elliot.recommender.gan.CFGAN.cfgan module¶

Module description:

class elliot.recommender.gan.CFGAN.cfgan.CFGAN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

CFGAN: A Generic Collaborative Filtering Framework based on Generative Adversarial Networks

For further details, please refer to the paper

Parameters

factors – Number of latent factor
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_gan – Adversarial regularization coefficient
g_epochs – Number of epochs to train the generator for each IRGAN step
d_epochs – Number of epochs to train the discriminator for each IRGAN step
s_zr – Sampling parameter of zero-reconstruction
s_pm – Sampling parameter of partial-masking

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  CFGAN:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    l_gan: 0.001
    g_epochs: 5
    d_epochs: 1
    s_zr: 0.001
    s_pm: 0.001

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.gan.CFGAN.cfgan_model module¶

Module description:

class elliot.recommender.gan.CFGAN.cfgan_model.CFGAN_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(predictions, train_mask, k=100)[source]¶

predict(start, stop, **kwargs)[source]¶

Generates output predictions for the input samples.

Computation is done in batches. This method is designed for performance in large scale inputs. For small amount of inputs that fit in one batch, directly using __call__ is recommended for faster execution, e.g., model(x), or model(x, training=False) if you have layers such as tf.keras.layers.BatchNormalization that behaves differently during inference. Also, note the fact that test loss is not affected by regularization layers like noise and dropout.

Parameters

x –
Input samples. It could be: - A Numpy array (or array-like), or a list of arrays

(in case the model has multiple inputs).
- A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).
- A tf.data dataset.
- A generator or keras.utils.Sequence instance.
A more detailed description of unpacking behavior for iterator types (Dataset, generator, Sequence) is given in the Unpacking behavior for iterator-like inputs section of Model.fit.
batch_size – Integer or None. Number of samples per batch. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of dataset, generators, or keras.utils.Sequence instances (since they generate batches).
verbose – Verbosity mode, 0 or 1.
steps – Total number of steps (batches of samples) before declaring the prediction round finished. Ignored with the default value of None. If x is a tf.data dataset and steps is None, predict will run until the input dataset is exhausted.
callbacks – List of keras.callbacks.Callback instances. List of callbacks to apply during prediction. See [callbacks](/api_docs/python/tf/keras/callbacks).
max_queue_size – Integer. Used for generator or keras.utils.Sequence input only. Maximum size for the generator queue. If unspecified, max_queue_size will default to 10.
workers – Integer. Used for generator or keras.utils.Sequence input only. Maximum number of processes to spin up when using process-based threading. If unspecified, workers will default to 1. If 0, will execute the generator on the main thread.
use_multiprocessing – Boolean. Used for generator or keras.utils.Sequence input only. If True, use process-based threading. If unspecified, use_multiprocessing will default to False. Note that because this implementation relies on multiprocessing, you should not pass non-picklable arguments to the generator as they can’t be passed easily to children processes.

See the discussion of Unpacking behavior for iterator-like inputs for Model.fit. Note that Model.predict uses the same interpretation rules as Model.fit and Model.evaluate, so inputs must be unambiguous for all three methods.

Returns

Numpy array(s) of predictions.

Raises

RuntimeError – If model.predict is wrapped in tf.function.
ValueError – In case of mismatch between the provided input data and the model’s expectations, or in case a stateful model receives a number of samples that is not a multiple of the batch size.

train_step(batch)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

class elliot.recommender.gan.CFGAN.cfgan_model.Discriminator(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

discriminate_fake_data(X)[source]¶

train_step(batch)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

class elliot.recommender.gan.CFGAN.cfgan_model.Generator(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

generate_fake_data(mask, C_u)[source]¶

infer(C_u)[source]¶

train_step(batch)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

Module contents¶

elliot.recommender.gan.IRGAN package¶

Submodules¶

elliot.recommender.gan.IRGAN.irgan module¶

Module description:

class elliot.recommender.gan.IRGAN.irgan.IRGAN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models

For further details, please refer to the paper

Parameters

factors – Number of latent factor
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_gan – Adversarial regularization coefficient
predict_model – Specification of the model to generate the recommendation (Generator/ Discriminator)
g_epochs – Number of epochs to train the generator for each IRGAN step
d_epochs – Number of epochs to train the discriminator for each IRGAN step
g_pretrain_epochs – Number of epochs to pre-train the generator
d_pretrain_epochs – Number of epochs to pre-train the discriminator
sample_lambda – Temperature Parameters

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  IRGAN:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.1
    l_b: 0.001
    l_gan: 0.001
    predict_model: generator
    g_epochs: 5
    d_epochs: 1
    g_pretrain_epochs: 10
    d_pretrain_epochs: 10
    sample_lambda: 0.2

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.gan.IRGAN.irgan_model module¶

Module description:

class elliot.recommender.gan.IRGAN.irgan_model.Discriminator(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

train_step(batch)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

class elliot.recommender.gan.IRGAN.irgan_model.Generator(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

train_step(batch)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

train_step_with_reward(batch)[source]¶

class elliot.recommender.gan.IRGAN.irgan_model.IRGAN_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_positions(predictions, train_mask, items, inner_test_user_true_mask)[source]¶

get_top_k(predictions, train_mask, k=100)[source]¶

pre_train_discriminator()[source]¶

pre_train_generator()[source]¶

predict(start, stop, **kwargs)[source]¶

Generates output predictions for the input samples.

Computation is done in batches. This method is designed for performance in large scale inputs. For small amount of inputs that fit in one batch, directly using __call__ is recommended for faster execution, e.g., model(x), or model(x, training=False) if you have layers such as tf.keras.layers.BatchNormalization that behaves differently during inference. Also, note the fact that test loss is not affected by regularization layers like noise and dropout.

Parameters

x –
Input samples. It could be: - A Numpy array (or array-like), or a list of arrays

(in case the model has multiple inputs).
- A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).
- A tf.data dataset.
- A generator or keras.utils.Sequence instance.
A more detailed description of unpacking behavior for iterator types (Dataset, generator, Sequence) is given in the Unpacking behavior for iterator-like inputs section of Model.fit.
batch_size – Integer or None. Number of samples per batch. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of dataset, generators, or keras.utils.Sequence instances (since they generate batches).
verbose – Verbosity mode, 0 or 1.
steps – Total number of steps (batches of samples) before declaring the prediction round finished. Ignored with the default value of None. If x is a tf.data dataset and steps is None, predict will run until the input dataset is exhausted.
callbacks – List of keras.callbacks.Callback instances. List of callbacks to apply during prediction. See [callbacks](/api_docs/python/tf/keras/callbacks).
max_queue_size – Integer. Used for generator or keras.utils.Sequence input only. Maximum size for the generator queue. If unspecified, max_queue_size will default to 10.
workers – Integer. Used for generator or keras.utils.Sequence input only. Maximum number of processes to spin up when using process-based threading. If unspecified, workers will default to 1. If 0, will execute the generator on the main thread.
use_multiprocessing – Boolean. Used for generator or keras.utils.Sequence input only. If True, use process-based threading. If unspecified, use_multiprocessing will default to False. Note that because this implementation relies on multiprocessing, you should not pass non-picklable arguments to the generator as they can’t be passed easily to children processes.

See the discussion of Unpacking behavior for iterator-like inputs for Model.fit. Note that Model.predict uses the same interpretation rules as Model.fit and Model.evaluate, so inputs must be unambiguous for all three methods.

Returns

Numpy array(s) of predictions.

Raises

RuntimeError – If model.predict is wrapped in tf.function.
ValueError – In case of mismatch between the provided input data and the model’s expectations, or in case a stateful model receives a number of samples that is not a multiple of the batch size.

train_step()[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

Module contents¶

elliot.recommender.graph_based package¶

Subpackages¶

elliot.recommender.graph_based.lightgcn package¶

Submodules¶

elliot.recommender.graph_based.lightgcn.LightGCN module¶

Module description:

class elliot.recommender.graph_based.lightgcn.LightGCN.LightGCN(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
l_w – Regularization coefficient
n_layers – Number of embedding propagation layers
n_fold – Number of folds to split the adjacency matrix into sub-matrices and ease the computation

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  LightGCN:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 64
    batch_size: 256
    l_w: 0.1
    n_layers: 1
    n_fold: 5

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.graph_based.lightgcn.LightGCN_model module¶

Module description:

class elliot.recommender.graph_based.lightgcn.LightGCN_model.LightGCNModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, **kwargs)[source]¶

Generates prediction for passed users and items indices

Parameters

inputs – user, item (batch)
Network in training mode or inference mode. (the) –

Returns

prediction and extracted model parameters

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(start, stop, **kwargs)[source]¶

train_step(batch)[source]¶

Apply a single training step on one batch.

Parameters: batch – batch used for the current train step
Returns: loss value at the current batch

Module contents¶

elliot.recommender.graph_based.ngcf package¶

Submodules¶

elliot.recommender.graph_based.ngcf.NGCF module¶

Module description:

class elliot.recommender.graph_based.ngcf.NGCF.NGCF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Graph Collaborative Filtering

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
l_w – Regularization coefficient
weight_size – Tuple with number of units for each embedding propagation layer
node_dropout – Tuple with dropout rate for each node
message_dropout – Tuple with dropout rate for each embedding propagation layer
n_fold – Number of folds to split the adjacency matrix into sub-matrices and ease the computation

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NGCF:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 64
    batch_size: 256
    l_w: 0.1
    weight_size: (64,)
    node_dropout: ()
    message_dropout: (0.1,)
    n_fold: 5

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.graph_based.ngcf.NGCF_model module¶

Module description:

class elliot.recommender.graph_based.ngcf.NGCF_model.NGCFModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, **kwargs)[source]¶

Generates prediction for passed users and items indices

Parameters

inputs – user, item (batch)
Network in training mode or inference mode. (the) –

Returns

prediction and extracted model parameters

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(start, stop, **kwargs)[source]¶

train_step(batch)[source]¶

Apply a single training step on one batch.

Parameters: batch – batch used for the current train step
Returns: loss value at the current batch

Module contents¶

elliot.recommender.knowledge_aware package¶

Subpackages¶

elliot.recommender.knowledge_aware.kaHFM package¶

Submodules¶

elliot.recommender.knowledge_aware.kaHFM.ka_hfm module¶

class elliot.recommender.knowledge_aware.kaHFM.ka_hfm.KaHFM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Knowledge-aware Hybrid Factorization Machines

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “How to Make Latent Factors Interpretable by Feeding Factorization Machines with Knowledge Graphs”, ISWC 2019 Best student Research Paper For further details, please refer to the paper

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “Semantic Interpretation of Top-N Recommendations”, IEEE TKDE 2020 For further details, please refer to the paper

Parameters

lr – learning rate (default: 0.05)
bias_regularization – Bias regularization (default: 0)
user_regularization – User regularization (default: 0.0025)
positive_item_regularization – regularization for positive (experienced) items (default: 0.0025)
negative_item_regularization – regularization for unknown items (default: 0.00025)
update_negative_item_factors – Boolean to update negative item factors (default: True)
update_users – Boolean to update user factors (default: True)
update_items – Boolean to update item factors (default: True)
update_bias – Boolean to update bias value (default: True)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  KaHFM:
    meta:
      hyper_max_evals: 20
      hyper_opt_alg: tpe
      validation_rate: 1
      verbose: True
      save_weights: True
      save_recs: True
      validation_metric: nDCG@10
    epochs: 100
    batch_size: -1
    lr: 0.05
    bias_regularization: 0
    user_regularization: 0.0025
    positive_item_regularization: 0.0025
    negative_item_regularization: 0.00025
    update_negative_item_factors: True
    update_users: True
    update_items: True
    update_bias: True

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

Get prediction on the user item pair.

Returns: A single float vaue.

restore_weights()[source]¶

train()[source]¶

train_step()[source]¶

update_factors(u: int, i: int, j: int)[source]¶

class elliot.recommender.knowledge_aware.kaHFM.ka_hfm.MF(ratings: Dict, map: Dict, tfidf: Dict, user_profiles: Dict, random: Any, *args)[source]¶

Bases: object

Simple Matrix Factorization class

get_factors()[source]¶

get_item_bias(item: int)[source]¶

get_item_factors(item: int)[source]¶

get_model_state()[source]¶

get_transactions()[source]¶

get_user_bias(user: int)[source]¶

get_user_factors(user: int)[source]¶

get_user_recs(user: int, k: int)[source]¶

get_user_recs_argpartition(user: int, k: int)[source]¶

initialize(loc: float = 0, scale: float = 0.1)[source]¶: This function initialize the data model :param loc: :param scale: :return:

property name¶

predict(user: int, item: int)[source]¶

set_item_bias(item: int, v: float)[source]¶

set_item_factors(item: int, v: float)[source]¶

set_model_state(saving_dict)[source]¶

set_user_bias(user: int, v: float)[source]¶

set_user_factors(user: int, v: float)[source]¶

elliot.recommender.knowledge_aware.kaHFM.tfidf_utils module¶

class elliot.recommender.knowledge_aware.kaHFM.tfidf_utils.TFIDF(map: Dict[int, List[int]])[source]¶

Bases: object

get_profiles(ratings: Dict[int, Dict[int, float]])[source]¶

tfidf()[source]¶

Module contents¶

elliot.recommender.knowledge_aware.kaHFM_batch package¶

Submodules¶

elliot.recommender.knowledge_aware.kaHFM_batch.kahfm_batch module¶

Module description:

class elliot.recommender.knowledge_aware.kaHFM_batch.kahfm_batch.KaHFMBatch(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Knowledge-aware Hybrid Factorization Machines (Tensorflow Batch Variant)

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “How to Make Latent Factors Interpretable by Feeding Factorization Machines with Knowledge Graphs”, ISWC 2019 Best student Research Paper For further details, please refer to the paper

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “Semantic Interpretation of Top-N Recommendations”, IEEE TKDE 2020 For further details, please refer to the paper

Parameters

lr – learning rate (default: 0.0001)
l_w – Weight regularization (default: 0.005)
l_b – Bias regularization (default: 0)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  KaHFMBatch:
    meta:
      hyper_max_evals: 20
      hyper_opt_alg: tpe
      validation_rate: 1
      verbose: True
      save_weights: True
      save_recs: True
      validation_metric: nDCG@10
    epochs: 100
    batch_size: -1
    lr: 0.0001
    l_w: 0.005
    l_b: 0

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.knowledge_aware.kaHFM_batch.kahfm_batch_model module¶

Module description:

class elliot.recommender.knowledge_aware.kaHFM_batch.kahfm_batch_model.KaHFM_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, **kwargs)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

predict_all()[source]¶

predict_batch(start, stop)[source]¶

train_step(batch)[source]¶

elliot.recommender.knowledge_aware.kaHFM_batch.tfidf_utils module¶

class elliot.recommender.knowledge_aware.kaHFM_batch.tfidf_utils.TFIDF(map: Dict[int, List[int]])[source]¶

Bases: object

get_profiles(ratings: Dict[int, Dict[int, float]])[source]¶

tfidf()[source]¶

Module contents¶

elliot.recommender.knowledge_aware.kahfm_embeddings package¶

Submodules¶

elliot.recommender.knowledge_aware.kahfm_embeddings.kahfm_embeddings module¶

Module description:

class elliot.recommender.knowledge_aware.kahfm_embeddings.kahfm_embeddings.KaHFMEmbeddings(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Knowledge-aware Hybrid Factorization Machines (Tensorflow Embedding-based Variant)

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “How to Make Latent Factors Interpretable by Feeding Factorization Machines with Knowledge Graphs”, ISWC 2019 Best student Research Paper For further details, please refer to the paper

Vito Walter Anelli and Tommaso Di Noia and Eugenio Di Sciascio and Azzurra Ragone and Joseph Trotta “Semantic Interpretation of Top-N Recommendations”, IEEE TKDE 2020 For further details, please refer to the paper

Parameters

lr – learning rate (default: 0.0001)
l_w – Weight regularization (default: 0.005)
l_b – Bias regularization (default: 0)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  KaHFMEmbeddings:
    meta:
      hyper_max_evals: 20
      hyper_opt_alg: tpe
      validation_rate: 1
      verbose: True
      save_weights: True
      save_recs: True
      validation_metric: nDCG@10
    epochs: 100
    batch_size: -1
    lr: 0.0001
    l_w: 0.005
    l_b: 0

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.knowledge_aware.kahfm_embeddings.kahfm_embeddings_model module¶

Module description:

class elliot.recommender.knowledge_aware.kahfm_embeddings.kahfm_embeddings_model.KaHFMEmbeddingsModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, **kwargs)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

predict_batch(start, stop)[source]¶

train_step(batch)[source]¶

elliot.recommender.knowledge_aware.kahfm_embeddings.tfidf_utils module¶

class elliot.recommender.knowledge_aware.kahfm_embeddings.tfidf_utils.TFIDF(map: Dict[int, List[int]])[source]¶

Bases: object

get_profiles(ratings: Dict[int, Dict[int, float]])[source]¶

tfidf()[source]¶

Module contents¶

elliot.recommender.latent_factor_models package¶

Subpackages¶

elliot.recommender.latent_factor_models.BPRMF package¶

Submodules¶

elliot.recommender.latent_factor_models.BPRMF.BPRMF module¶

Module description:

class elliot.recommender.latent_factor_models.BPRMF.BPRMF.BPRMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Bayesian Personalized Ranking with Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
bias_regularization – Regularization coefficient for the bias
user_regularization – Regularization coefficient for user latent factors
positive_item_regularization – Regularization coefficient for positive item latent factors
negative_item_regularization – Regularization coefficient for negative item latent factors
update_negative_item_factors –
update_users –
update_items –
update_bias –

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  BPRMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    bias_regularization: 0
    user_regularization: 0.0025
    positive_item_regularization: 0.0025
    negative_item_regularization: 0.0025
    update_negative_item_factors: True
    update_users: True
    update_items: True
    update_bias: True

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

Get prediction on the user item pair.

Returns: A single float vaue.

restore_weights()[source]¶

train()[source]¶

train_step()[source]¶

update_factors(u: int, i: int, j: int)[source]¶

class elliot.recommender.latent_factor_models.BPRMF.BPRMF.MF(F, ratings, random, *args)[source]¶

Bases: object

get_item_bias(item: int)[source]¶

get_item_factors(item: int)[source]¶

get_model_state()[source]¶

get_transactions()[source]¶

get_user_bias(user: int)[source]¶

get_user_factors(user: int)[source]¶

get_user_recs(user, k)[source]¶

get_user_recs_argpartition(user: int, k: int)[source]¶

initialize(loc: float = 0, scale: float = 0.1)[source]¶: This function initialize the data model :param loc: :param scale: :return:

property name¶

predict(user, item)[source]¶

set_item_bias(item: int, v: float)[source]¶

set_item_factors(item: int, v: float)[source]¶

set_model_state(saving_dict)[source]¶

set_user_bias(user: int, v: float)[source]¶

set_user_factors(user: int, v: float)[source]¶

Module contents¶

Module description:

elliot.recommender.latent_factor_models.BPRMF_batch package¶

Submodules¶

elliot.recommender.latent_factor_models.BPRMF_batch.BPRMF_batch module¶

Module description:

class elliot.recommender.latent_factor_models.BPRMF_batch.BPRMF_batch.BPRMF_batch(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Batch Bayesian Personalized Ranking with Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
l_w – Regularization coefficient for latent factors
l_b – Regularization coefficient for bias

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  BPRMF_batch:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.1
    l_b: 0.001

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.BPRMF_batch.BPRMF_batch_model module¶

Module description:

class elliot.recommender.latent_factor_models.BPRMF_batch.BPRMF_batch_model.BPRMF_batch_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_positions(predictions, train_mask, items, inner_test_user_true_mask)[source]¶

get_top_k(predictions, train_mask, k=100)[source]¶

predict(start, stop, **kwargs)[source]¶

train_step(batch)[source]¶

Module contents¶

Module description:

elliot.recommender.latent_factor_models.BPRSlim package¶

Submodules¶

elliot.recommender.latent_factor_models.BPRSlim.bprslim module¶

Module description:

class elliot.recommender.latent_factor_models.BPRSlim.bprslim.BPRSlim(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

BPR Sparse Linear Methods

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
lj_reg – Regularization coefficient for positive items
li_reg – Regularization coefficient for negative items

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  AMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    lj_reg: 0.001
    li_reg: 0.1

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

Get prediction on the user item pair.

Returns: A single float vaue.

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.BPRSlim.bprslim_model module¶

Module description:

class elliot.recommender.latent_factor_models.BPRSlim.bprslim_model.BPRSlimModel(data, num_users, num_items, lr, lj_reg, li_reg, sampler, random_seed=42)[source]¶

Bases: object

get_model_state()[source]¶

get_user_recs(user, k=100)[source]¶

predict(u, i)[source]¶

set_model_state(saving_dict)[source]¶

train_step(batch)[source]¶

Module contents¶

Module description:

elliot.recommender.latent_factor_models.CML package¶

Submodules¶

elliot.recommender.latent_factor_models.CML.CML module¶

Module description:

class elliot.recommender.latent_factor_models.CML.CML.CML(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Collaborative Metric Learning

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
l_w – Regularization coefficient for latent factors
l_b – Regularization coefficient for bias
margin – Safety margin size

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  CML:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.001
    l_b: 0.001
    margin: 0.5

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.CML.CML_model module¶

Module description:

class elliot.recommender.latent_factor_models.CML.CML_model.CML_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_positions(predictions, train_mask, items, inner_test_user_true_mask)[source]¶

get_top_k(predictions, train_mask, k=100)[source]¶

predict(start, stop, **kwargs)[source]¶

train_step(batch)[source]¶

class elliot.recommender.latent_factor_models.CML.CML_model.LatentFactor(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.layers.embeddings.Embedding

censor(censor_id)[source]¶

Module contents¶

Module description:

elliot.recommender.latent_factor_models.FFM package¶

Submodules¶

elliot.recommender.latent_factor_models.FFM.field_aware_factorization_machine module¶

Module description:

class elliot.recommender.latent_factor_models.FFM.field_aware_factorization_machine.FFM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Field-aware Factorization Machines

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  FFM:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.FFM.field_aware_factorization_machine_model module¶

Module description:

class elliot.recommender.latent_factor_models.FFM.field_aware_factorization_machine_model.FieldAwareFactorizationMachineModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.latent_factor_models.FISM package¶

Submodules¶

elliot.recommender.latent_factor_models.FISM.FISM module¶

Module description:

class elliot.recommender.latent_factor_models.FISM.FISM.FISM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

FISM: Factored Item Similarity Models

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
l_w – Regularization coefficient for latent factors
l_b – Regularization coefficient for bias
alpha – Alpha parameter (a value between 0 and 1)
neg_ratio –

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  FISM:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    l_w: 0.001
    l_b: 0.001
    alpha: 0.5
    neg_ratio:

get_recommendations(k: int = 100, auc_compute: bool = False)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.FISM.FISM_model module¶

Module description:

class elliot.recommender.latent_factor_models.FISM.FISM_model.FISM_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

batch_predict(user_start, user_stop, **kwargs)[source]¶

call(inputs, training=None)[source]¶

create_history_item_matrix()[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_positions(predictions, train_mask, items, inner_test_user_true_mask)[source]¶

get_top_k(predictions, train_mask, k=100)[source]¶

predict(user, **kwargs)[source]¶

train_step(batch)[source]¶

class elliot.recommender.latent_factor_models.FISM.FISM_model.LatentFactor(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.layers.embeddings.Embedding

censor(censor_id)[source]¶

Module contents¶

Module description:

elliot.recommender.latent_factor_models.FM package¶

Submodules¶

elliot.recommender.latent_factor_models.FM.factorization_machine module¶

Module description:

class elliot.recommender.latent_factor_models.FM.factorization_machine.FM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Factorization Machines

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  FM:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.FM.factorization_machine_model module¶

Module description:

class elliot.recommender.latent_factor_models.FM.factorization_machine_model.FactorizationMachineModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.latent_factor_models.FunkSVD package¶

Submodules¶

elliot.recommender.latent_factor_models.FunkSVD.funk_svd module¶

Module description:

class elliot.recommender.latent_factor_models.FunkSVD.funk_svd.FunkSVD(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
reg_w – Regularization coefficient for latent factors
reg_b – Regularization coefficient for bias

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  FunkSVD:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg_w: 0.1
    reg_b: 0.001

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.FunkSVD.funk_svd_model module¶

Module description:

class elliot.recommender.latent_factor_models.FunkSVD.funk_svd_model.FunkSVDModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.latent_factor_models.LogisticMF package¶

Submodules¶

elliot.recommender.latent_factor_models.LogisticMF.logistic_matrix_factorization module¶

Module description:

class elliot.recommender.latent_factor_models.LogisticMF.logistic_matrix_factorization.LogisticMatrixFactorization(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Logistic Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of factors of feature embeddings
lr – Learning rate
reg – Regularization coefficient
alpha – Parameter for confidence estimation

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  LogisticMatrixFactorization:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1
    alpha: 0.5

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.LogisticMF.logistic_matrix_factorization_model module¶

Module description:

class elliot.recommender.latent_factor_models.LogisticMF.logistic_matrix_factorization_model.LogisticMatrixFactorizationModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_top_k(preds, train_mask, k=100)[source]¶

predict_batch(start, stop, **kwargs)[source]¶

set_update_user(update_user)[source]¶

train_step(batch)[source]¶

Module contents¶

elliot.recommender.latent_factor_models.MF package¶

Submodules¶

elliot.recommender.latent_factor_models.MF.matrix_factorization module¶

Module description:

class elliot.recommender.latent_factor_models.MF.matrix_factorization.MF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  MF:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.MF.matrix_factorization_model module¶

Module description:

class elliot.recommender.latent_factor_models.MF.matrix_factorization_model.MatrixFactorizationModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.latent_factor_models.NonNegMF package¶

Submodules¶

elliot.recommender.latent_factor_models.NonNegMF.non_negative_matrix_factorization module¶

Module description:

class elliot.recommender.latent_factor_models.NonNegMF.non_negative_matrix_factorization.NonNegMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Non-Negative Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NonNegMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    lr: 0.001
    reg: 0.1

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

Get prediction on the user item pair.

Returns: A single float vaue.

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.NonNegMF.non_negative_matrix_factorization_model module¶

Module description:

class elliot.recommender.latent_factor_models.NonNegMF.non_negative_matrix_factorization_model.NonNegMFModel(data, num_users, num_items, global_mean, embed_mf_size, lambda_weights, learning_rate=0.01, random_seed=42)[source]¶

Bases: object

get_model_state()[source]¶

get_user_recs(user, k=100)[source]¶

predict(user, item)[source]¶

set_model_state(saving_dict)[source]¶

train_step()[source]¶

Module contents¶

Module description:

elliot.recommender.latent_factor_models.PMF package¶

Submodules¶

elliot.recommender.latent_factor_models.PMF.probabilistic_matrix_factorization module¶

Module description:

Mnih, Andriy, and Russ R. Salakhutdinov. “Probabilistic matrix factorization.” Advances in neural information processing systems 20 (2007)

class elliot.recommender.latent_factor_models.PMF.probabilistic_matrix_factorization.PMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Probabilistic Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
reg – Regularization coefficient
gaussian_variance – Variance of the Gaussian distribution

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  PMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 50
    lr: 0.001
    reg: 0.0025
    gaussian_variance: 0.1

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.PMF.probabilistic_matrix_factorization_model module¶

Module description:

Mnih, Andriy, and Russ R. Salakhutdinov. “Probabilistic matrix factorization.” Advances in neural information processing systems 20 (2007)

class elliot.recommender.latent_factor_models.PMF.probabilistic_matrix_factorization_model.ProbabilisticMatrixFactorizationModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

dot_prod(layer_0, layer_1)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.latent_factor_models.PureSVD package¶

Submodules¶

elliot.recommender.latent_factor_models.PureSVD.pure_svd module¶

Module description:

class elliot.recommender.latent_factor_models.PureSVD.pure_svd.PureSVD(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

For further details, please refer to the paper

Parameters

factors – Number of latent factors
seed – Random seed

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  PureSVD:
    meta:
      save_recs: True
    epochs: 10
    factors: 10
    seed: 42

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

Get prediction on the user item pair.

Returns: A single float vaue.

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.PureSVD.pure_svd_model module¶

Module description:

class elliot.recommender.latent_factor_models.PureSVD.pure_svd_model.PureSVDModel(factors, data, random_seed)[source]¶

Bases: object

Simple Matrix Factorization class

get_model_state()[source]¶

get_user_recs(user, k=100)[source]¶

predict(user, item)[source]¶

set_model_state(saving_dict)[source]¶

train_step()[source]¶

Module contents¶

elliot.recommender.latent_factor_models.SVDpp package¶

Submodules¶

elliot.recommender.latent_factor_models.SVDpp.svdpp module¶

Module description:

class elliot.recommender.latent_factor_models.SVDpp.svdpp.SVDpp(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

SVD++

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
reg_w – Regularization coefficient for latent factors
reg_b – Regularization coefficient for bias

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  SVDpp:
    meta:
      save_recs: True
    epochs: 10
    factors: 50
    lr: 0.001
    reg_w: 0.1
    reg_b: 0.001

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.SVDpp.svdpp_model module¶

Module description:

class elliot.recommender.latent_factor_models.SVDpp.svdpp_model.SVDppModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.latent_factor_models.Slim package¶

Submodules¶

elliot.recommender.latent_factor_models.Slim.slim module¶

Module description:

class elliot.recommender.latent_factor_models.Slim.slim.Slim(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Sparse Linear Methods

For further details, please refer to the paper

Parameters

l1_ratio –
alpha –

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  Slim:
    meta:
      save_recs: True
    epochs: 10
    l1_ratio: 0.001
    alpha: 0.001

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

Get prediction on the user item pair.

Returns: A single float vaue.

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.Slim.slim_model module¶

Module description:

class elliot.recommender.latent_factor_models.Slim.slim_model.SlimModel(data, num_users, num_items, l1_ratio, alpha, epochs)[source]¶

Bases: object

get_model_state()[source]¶

get_user_recs(user, k=100)[source]¶

predict(u, i)[source]¶

set_model_state(saving_dict)[source]¶

train(verbose)[source]¶

Module contents¶

Module description:

elliot.recommender.latent_factor_models.WRMF package¶

Submodules¶

elliot.recommender.latent_factor_models.WRMF.wrmf module¶

Module description:

class elliot.recommender.latent_factor_models.WRMF.wrmf.WRMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Weighted XXX Matrix Factorization

For further details, please refer to the paper

Parameters

factors – Number of latent factors
lr – Learning rate
alpha –
reg – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  WRMF:
    meta:
      save_recs: True
    epochs: 10
    factors: 50
    alpha: 1
    reg: 0.1

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

Get prediction on the user item pair.

Returns: A single float vaue.

restore_weights()[source]¶

train()[source]¶

elliot.recommender.latent_factor_models.WRMF.wrmf_model module¶

Module description:

class elliot.recommender.latent_factor_models.WRMF.wrmf_model.WRMFModel(factors, data, random, alpha, reg)[source]¶

Bases: object

Simple Matrix Factorization class

get_model_state()[source]¶

get_user_recs(user, k=100)[source]¶

predict(user, item)[source]¶

set_model_state(saving_dict)[source]¶

train_step()[source]¶

Module contents¶

Module description:

elliot.recommender.neural package¶

Subpackages¶

elliot.recommender.neural.ConvMF package¶

Submodules¶

elliot.recommender.neural.ConvMF.convolutional_matrix_factorization module¶

Module description:

class elliot.recommender.neural.ConvMF.convolutional_matrix_factorization.ConvMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Convolutional Matrix Factorization for Document Context-Aware Recommendation

For further details, please refer to the paper

Parameters

embedding_size – Embedding dimension
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
cnn_channels – List of channels
cnn_kernels – List of kernels
cnn_strides – List of strides
dropout_prob – Dropout probability applied on the convolutional layers

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ConvMF:
    meta:
      save_recs: True
    epochs: 10
    embedding_size: 100
    lr: 0.001
    l_w: 0.005
    l_b: 0.0005
    cnn_channels: (1, 32, 32)
    cnn_kernels: (2,2)
    cnn_strides: (2,2)
    dropout_prob: 0

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.neural.ConvMF.convolutional_matrix_factorization_model module¶

Module description:

class elliot.recommender.neural.ConvMF.convolutional_matrix_factorization_model.ConvMatrixFactorizationModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=False, **kwargs)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

class elliot.recommender.neural.ConvMF.convolutional_matrix_factorization_model.ConvolutionalComponent(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, **kwargs)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

class elliot.recommender.neural.ConvMF.convolutional_matrix_factorization_model.MLPComponent(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=False, **kwargs)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

Module contents¶

elliot.recommender.neural.ConvNeuMF package¶

Submodules¶

elliot.recommender.neural.ConvNeuMF.convolutional_neural_matrix_factorization module¶

Module description:

class elliot.recommender.neural.ConvNeuMF.convolutional_neural_matrix_factorization.ConvNeuMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Outer Product-based Neural Collaborative Filtering

For further details, please refer to the paper

Parameters

embedding_size – Embedding dimension
lr – Learning rate
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
cnn_channels – List of channels
cnn_kernels – List of kernels
cnn_strides – List of strides
dropout_prob – Dropout probability applied on the convolutional layers

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ConvNeuMF:
    meta:
      save_recs: True
    epochs: 10
    embedding_size: 100
    lr: 0.001
    l_w: 0.005
    l_b: 0.0005
    cnn_channels: (1, 32, 32)
    cnn_kernels: (2,2)
    cnn_strides: (2,2)
    dropout_prob: 0

get_recommendations(k: int = 100)[source]¶

property name¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.neural.ConvNeuMF.convolutional_neural_matrix_factorization_model module¶

Module description:

class elliot.recommender.neural.ConvNeuMF.convolutional_neural_matrix_factorization_model.ConvNeuralMatrixFactorizationModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=False, **kwargs)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

class elliot.recommender.neural.ConvNeuMF.convolutional_neural_matrix_factorization_model.ConvolutionalComponent(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, **kwargs)[source]¶

class elliot.recommender.neural.ConvNeuMF.convolutional_neural_matrix_factorization_model.MLPComponent(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=False, **kwargs)[source]¶

Module contents¶

elliot.recommender.neural.DMF package¶

Submodules¶

elliot.recommender.neural.DMF.deep_matrix_factorization module¶

Module description:

class elliot.recommender.neural.DMF.deep_matrix_factorization.DMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Deep Matrix Factorization Models for Recommender Systems.

For further details, please refer to the paper

Parameters

lr – Learning rate
reg – Regularization coefficient
user_mlp – List of units for each layer
item_mlp – List of activation functions
similarity – Number of factors dimension

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  DMF:
    meta:
      save_recs: True
    epochs: 10
    lr: 0.0001
    reg: 0.001
    user_mlp: (64,32)
    item_mlp: (64,32)
    similarity: cosine

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.neural.DMF.deep_matrix_factorization_model module¶

Module description:

class elliot.recommender.neural.DMF.deep_matrix_factorization_model.DeepMatrixFactorizationModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

cosine(layer_0, layer_1)[source]¶

dot_prod(layer_0, layer_1)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.neural.DeepFM package¶

Submodules¶

elliot.recommender.neural.DeepFM.deep_fm module¶

Module description:

class elliot.recommender.neural.DeepFM.deep_fm.DeepFM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

DeepFM: A Factorization-Machine based Neural Network for CTR Prediction

For further details, please refer to the paper

Parameters

factors – Number of factors dimension
lr – Learning rate
l_w – Regularization coefficient
hidden_neurons – List of units for each layer
hidden_activations – List of activation functions

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  DeepFM:
    meta:
      save_recs: True
    epochs: 10
    factors: 100
    lr: 0.001
    l_w: 0.0001
    hidden_neurons: (64,32)
    hidden_activations: ('relu','relu')

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.neural.DeepFM.deep_fm_model module¶

Module description:

class elliot.recommender.neural.DeepFM.deep_fm_model.DeepFMModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.neural.GeneralizedMF package¶

Submodules¶

elliot.recommender.neural.GeneralizedMF.generalized_matrix_factorization module¶

Module description:

class elliot.recommender.neural.GeneralizedMF.generalized_matrix_factorization.GMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Collaborative Filtering

For further details, please refer to the paper

Parameters

mf_factors – Number of latent factors
lr – Learning rate
is_edge_weight_train – Whether the training uses edge weighting

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  GMF:
    meta:
      save_recs: True
    epochs: 10
    mf_factors: 10
    lr: 0.001
    is_edge_weight_train: True

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

train()[source]¶

elliot.recommender.neural.GeneralizedMF.generalized_matrix_factorization_model module¶

Module description:

class elliot.recommender.neural.GeneralizedMF.generalized_matrix_factorization_model.GeneralizedMatrixFactorizationModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

train_step(batch)[source]¶

Module contents¶

elliot.recommender.neural.ItemAutoRec package¶

Submodules¶

elliot.recommender.neural.ItemAutoRec.itemautorec module¶

Module description:

class elliot.recommender.neural.ItemAutoRec.itemautorec.ItemAutoRec(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

AutoRec: Autoencoders Meet Collaborative Filtering (Item-based)

For further details, please refer to the paper

Parameters

hidden_neuron – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ItemAutoRec:
    meta:
      save_recs: True
    epochs: 10
    hidden_neuron: 500
    lr: 0.0001
    l_w: 0.001

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.neural.ItemAutoRec.itemautorec_model module¶

Module description:

class elliot.recommender.neural.ItemAutoRec.itemautorec_model.Decoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

call(inputs, **kwargs)[source]¶

class elliot.recommender.neural.ItemAutoRec.itemautorec_model.Encoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

call(inputs, training=None)[source]¶

class elliot.recommender.neural.ItemAutoRec.itemautorec_model.ItemAutoRecModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, **kwargs)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix. [Is Inverted]

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

Module description:

elliot.recommender.neural.NAIS package¶

Submodules¶

elliot.recommender.neural.NAIS.nais module¶

Module description:

class elliot.recommender.neural.NAIS.nais.NAIS(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

NAIS: Neural Attentive Item Similarity Model for Recommendation

For further details, please refer to the paper

Parameters

factors – Number of latent factors
algorithm – Type of user-item factor operation (‘product’, ‘concat’)
weight_size – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient
l_b – Bias regularization coefficient
alpha – Attention factor
beta – Smoothing exponent
neg_ratio – Ratio of negative sampled items, e.g., 0 = no items, 1 = all un-rated items

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NAIS:
    meta:
      save_recs: True
    factors: 100
    algorithm: concat
    weight_size: 32
    lr: 0.001
    l_w: 0.001
    l_b: 0.001
    alpha: 0.5
    beta: 0.5
    neg_ratio: 0.5

get_recommendations(k: int = 100, auc_compute: bool = False)[source]¶

property name¶

train()[source]¶

elliot.recommender.neural.NAIS.nais_model module¶

Module description:

class elliot.recommender.neural.NAIS.nais_model.LatentFactor(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.layers.embeddings.Embedding

censor(censor_id)[source]¶

class elliot.recommender.neural.NAIS.nais_model.NAIS_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

attention(user_history, target)[source]¶

batch_attention(user_history, target)[source]¶

batch_predict(user_start, user_stop)[source]¶

batch_softmax(logits, item_num, similarity, user_bias, item_bias)[source]¶

call(inputs, training=None)[source]¶

create_history_item_matrix()[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_positions(predictions, train_mask, items, inner_test_user_true_mask)[source]¶

get_top_k(predictions, train_mask, k=100)[source]¶

predict(user, **kwargs)[source]¶

softmax(logits, item_num, similarity, user_bias, item_bias, batch_mask_mat=None)[source]¶

train_step(batch)[source]¶

Module contents¶

elliot.recommender.neural.NFM package¶

Submodules¶

elliot.recommender.neural.NFM.neural_fm module¶

Module description:

class elliot.recommender.neural.NFM.neural_fm.NFM(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Factorization Machines for Sparse Predictive Analytics

For further details, please refer to the paper

Parameters

factors – Number of factors dimension
lr – Learning rate
l_w – Regularization coefficient
hidden_neurons – List of units for each layer
hidden_activations – List of activation functions

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NFM:
    meta:
      save_recs: True
    epochs: 10
    factors: 100
    lr: 0.001
    l_w: 0.0001
    hidden_neurons: (64,32)
    hidden_activations: ('relu','relu')

get_recommendations(k: int = 100)[source]¶

property name¶

predict(u: int, i: int)[source]¶

restore_weights()[source]¶

train()[source]¶

elliot.recommender.neural.NFM.neural_fm_model module¶

Module description:

class elliot.recommender.neural.NFM.neural_fm_model.NeuralFactorizationMachineModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.neural.NPR package¶

Submodules¶

elliot.recommender.neural.NPR.neural_personalized_ranking module¶

Module description:

class elliot.recommender.neural.NPR.neural_personalized_ranking.NPR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Personalized Ranking for Image Recommendation (Model without visual features)

For further details, please refer to the paper

Parameters

mf_factors – Number of MF latent factors
mlp_hidden_size – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient
dropout – Dropout rate

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NPR:
    meta:
      save_recs: True
    epochs: 10
    mf_factors: 100
    mlp_hidden_size:  (64,32)
    lr: 0.001
    l_w: 0.001
    dropout: 0.45

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.neural.NPR.neural_personalized_ranking_model module¶

Module description:

class elliot.recommender.neural.NPR.neural_personalized_ranking_model.NPRModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

Module contents¶

elliot.recommender.neural.NeuMF package¶

Submodules¶

elliot.recommender.neural.NeuMF.neural_matrix_factorization module¶

Module description:

class elliot.recommender.neural.NeuMF.neural_matrix_factorization.NeuMF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Neural Collaborative Filtering

For further details, please refer to the paper

Parameters

mf_factors – Number of MF latent factors
mlp_factors – Number of MLP latent factors
mlp_hidden_size – List of units for each layer
lr – Learning rate
dropout – Dropout rate
is_mf_train – Whether to train the MF embeddings
is_mlp_train – Whether to train the MLP layers

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NeuMF:
    meta:
      save_recs: True
    epochs: 10
    mf_factors: 10
    mlp_factors: 10
    mlp_hidden_size: (64,32)
    lr: 0.001
    dropout: 0.0
    is_mf_train: True
    is_mlp_train: True

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.neural.NeuMF.neural_matrix_factorization_model module¶

Module description:

class elliot.recommender.neural.NeuMF.neural_matrix_factorization_model.NeuralMatrixFactorizationModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

elliot.recommender.neural.UserAutoRec package¶

Submodules¶

elliot.recommender.neural.UserAutoRec.userautorec module¶

Module description:

class elliot.recommender.neural.UserAutoRec.userautorec.UserAutoRec(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

AutoRec: Autoencoders Meet Collaborative Filtering (User-based)

For further details, please refer to the paper

Parameters

hidden_neuron – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  UserAutoRec:
    meta:
      save_recs: True
    epochs: 10
    hidden_neuron: 500
    lr: 0.0001
    l_w: 0.001

property name¶

train()[source]¶

elliot.recommender.neural.UserAutoRec.userautorec_model module¶

Module description:

class elliot.recommender.neural.UserAutoRec.userautorec_model.Decoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

call(inputs, **kwargs)[source]¶

class elliot.recommender.neural.UserAutoRec.userautorec_model.Encoder(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.base_layer.Layer

call(inputs, training=None)[source]¶

class elliot.recommender.neural.UserAutoRec.userautorec_model.UserAutoRecModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, **kwargs)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

Module description:

elliot.recommender.neural.WideAndDeep package¶

Submodules¶

elliot.recommender.neural.WideAndDeep.wide_and_deep module¶

Module description:

class elliot.recommender.neural.WideAndDeep.wide_and_deep.WideAndDeep(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Wide & Deep Learning for Recommender Systems

(For now, available with knowledge-aware features)

For further details, please refer to the paper

Parameters

factors – Number of latent factors
mlp_hidden_size – List of units for each layer
lr – Learning rate
l_w – Regularization coefficient
l_b – Bias Regularization Coefficient
dropout_prob – Dropout rate

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  NPR:
    meta:
      save_recs: True
    epochs: 10
    factors: 50
    mlp_hidden_size: (32, 32, 1)
    lr: 0.001
    l_w: 0.005
    l_b: 0.0005
    dropout_prob: 0.0

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.neural.WideAndDeep.wide_and_deep.build_sparse_features(data)[source]¶

elliot.recommender.neural.WideAndDeep.wide_and_deep_model module¶

Module description:

class elliot.recommender.neural.WideAndDeep.wide_and_deep_model.WideAndDeepModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=False, **kwargs)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

get_sparse(u, i)[source]¶

get_top_k(preds, train_mask, k=100)[source]¶

get_user_recs(user, k=100)[source]¶

predict(user, **kwargs)[source]¶

Generates output predictions for the input samples.

Computation is done in batches. This method is designed for performance in large scale inputs. For small amount of inputs that fit in one batch, directly using __call__ is recommended for faster execution, e.g., model(x), or model(x, training=False) if you have layers such as tf.keras.layers.BatchNormalization that behaves differently during inference. Also, note the fact that test loss is not affected by regularization layers like noise and dropout.

Parameters

x –
Input samples. It could be: - A Numpy array (or array-like), or a list of arrays

(in case the model has multiple inputs).
- A TensorFlow tensor, or a list of tensors (in case the model has multiple inputs).
- A tf.data dataset.
- A generator or keras.utils.Sequence instance.
A more detailed description of unpacking behavior for iterator types (Dataset, generator, Sequence) is given in the Unpacking behavior for iterator-like inputs section of Model.fit.
batch_size – Integer or None. Number of samples per batch. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of dataset, generators, or keras.utils.Sequence instances (since they generate batches).
verbose – Verbosity mode, 0 or 1.
steps – Total number of steps (batches of samples) before declaring the prediction round finished. Ignored with the default value of None. If x is a tf.data dataset and steps is None, predict will run until the input dataset is exhausted.
callbacks – List of keras.callbacks.Callback instances. List of callbacks to apply during prediction. See [callbacks](/api_docs/python/tf/keras/callbacks).
max_queue_size – Integer. Used for generator or keras.utils.Sequence input only. Maximum size for the generator queue. If unspecified, max_queue_size will default to 10.
workers – Integer. Used for generator or keras.utils.Sequence input only. Maximum number of processes to spin up when using process-based threading. If unspecified, workers will default to 1. If 0, will execute the generator on the main thread.
use_multiprocessing – Boolean. Used for generator or keras.utils.Sequence input only. If True, use process-based threading. If unspecified, use_multiprocessing will default to False. Note that because this implementation relies on multiprocessing, you should not pass non-picklable arguments to the generator as they can’t be passed easily to children processes.

See the discussion of Unpacking behavior for iterator-like inputs for Model.fit. Note that Model.predict uses the same interpretation rules as Model.fit and Model.evaluate, so inputs must be unambiguous for all three methods.

Returns

Numpy array(s) of predictions.

Raises

RuntimeError – If model.predict is wrapped in tf.function.
ValueError – In case of mismatch between the provided input data and the model’s expectations, or in case a stateful model receives a number of samples that is not a multiple of the batch size.

train_step(batch)[source]¶

The logic for one training step.

This method can be overridden to support custom training logic. This method is called by Model.make_train_function.

This method should contain the mathemetical logic for one step of training. This typically includes the forward pass, loss calculation, backpropagation, and metric updates.

Configuration details for how this logic is run (e.g. tf.function and tf.distribute.Strategy settings), should be left to Model.make_train_function, which can also be overridden.

Parameters: data – A nested structure of `Tensor`s.
Returns: A dict containing values that will be passed to tf.keras.callbacks.CallbackList.on_train_batch_end. Typically, the values of the Model’s metrics are returned. Example: {‘loss’: 0.2, ‘accuracy’: 0.7}.

Module contents¶

elliot.recommender.unpersonalized package¶

Subpackages¶

elliot.recommender.unpersonalized.most_popular package¶

Submodules¶

elliot.recommender.unpersonalized.most_popular.most_popular module¶

Created on April 4, 2020 Tensorflow 2.1.0 implementation of APR. @author Anonymized

class elliot.recommender.unpersonalized.most_popular.most_popular.MostPop(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

get_recommendations(top_k)[source]¶

property name¶

train()[source]¶

Module contents¶

elliot.recommender.unpersonalized.random_recommender package¶

Submodules¶

elliot.recommender.unpersonalized.random_recommender.Random module¶

Created on April 4, 2020 Tensorflow 2.1.0 implementation of APR. @author Anonymized

class elliot.recommender.unpersonalized.random_recommender.Random.Random(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

get_recommendations(top_k)[source]¶

property name¶

train()[source]¶

Module contents¶

elliot.recommender.visual_recommenders package¶

Subpackages¶

elliot.recommender.visual_recommenders.ACF package¶

Submodules¶

elliot.recommender.visual_recommenders.ACF.ACF module¶

Module description:

class elliot.recommender.visual_recommenders.ACF.ACF.ACF(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Attentive Collaborative Filtering: Multimedia Recommendation with Item- and Component-Level Attention

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
l_w – Regularization coefficient
layers_component – Tuple with number of units for each attentive layer (component-level)
layers_item – Tuple with number of units for each attentive layer (item-level)

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  ACF:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 100
    batch_size: 128
    l_w: 0.000025
    layers_component: (64, 1)
    layers_item: (64, 1)

property name¶

train()[source]¶

elliot.recommender.visual_recommenders.ACF.ACF_model module¶

Module description:

class elliot.recommender.visual_recommenders.ACF.ACF_model.ACF_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(start, stop)[source]¶

train_step(batch)[source]¶

Module contents¶

Module description:

elliot.recommender.visual_recommenders.DVBPR package¶

Submodules¶

elliot.recommender.visual_recommenders.DVBPR.DVBPR module¶

Module description:

class elliot.recommender.visual_recommenders.DVBPR.DVBPR.DVBPR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Visually-Aware Fashion Recommendation and Design with Generative Image Models

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
lambda_1 – Regularization coefficient
lambda_2 – CNN regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  DVBPR:
    meta:
      save_recs: True
    lr: 0.0001
    epochs: 50
    factors: 100
    batch_size: 128
    lambda_1: 0.0001
    lambda_2: 1.0

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.visual_recommenders.DVBPR.DVBPR_model module¶

Module description:

class elliot.recommender.visual_recommenders.DVBPR.DVBPR_model.DVBPR_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_config()[source]¶

get_top_k(preds, train_mask, k=100)[source]¶

predict_batch(start, stop, phi)[source]¶

train_step(batch)[source]¶

elliot.recommender.visual_recommenders.DVBPR.FeatureExtractor module¶

class elliot.recommender.visual_recommenders.DVBPR.FeatureExtractor.FeatureExtractor(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model, abc.ABC

call(inputs, training=None, mask=None)[source]¶

Calls the model on new inputs.

In this case call just reapplies all ops in the graph to the new inputs (e.g. build a new computational graph from the provided inputs).

Parameters

inputs – A tensor or list of tensors.
training – Boolean or boolean scalar tensor, indicating whether to run the Network in training mode or inference mode.
mask – A mask or list of masks. A mask can be either a tensor or None (no mask).

Returns

A tensor if there is a single output, or a list of tensors if there are more than one outputs.

Module contents¶

Module description:

elliot.recommender.visual_recommenders.DeepStyle package¶

Submodules¶

elliot.recommender.visual_recommenders.DeepStyle.DeepStyle module¶

Module description:

class elliot.recommender.visual_recommenders.DeepStyle.DeepStyle.DeepStyle(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

DeepStyle: Learning User Preferences for Visual Recommendation

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
batch_size – Batch size
l_w – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  DeepStyle:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 100
    batch_size: 128
    l_w: 0.000025

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.visual_recommenders.DeepStyle.DeepStyle_model module¶

Module description:

class elliot.recommender.visual_recommenders.DeepStyle.DeepStyle_model.DeepStyle_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None)[source]¶

get_config()[source]¶

get_top_k(preds, train_mask, k=100)[source]¶

predict(start, stop, training=False)[source]¶

train_step(batch)[source]¶

Module contents¶

Module description:

elliot.recommender.visual_recommenders.VBPR package¶

Submodules¶

elliot.recommender.visual_recommenders.VBPR.VBPR module¶

Module description:

class elliot.recommender.visual_recommenders.VBPR.VBPR.VBPR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
factors – Number of latent factors
factors_d – Dimension of visual factors
batch_size – Batch size
l_w – Regularization coefficient
l_b – Regularization coefficient of bias
l_e – Regularization coefficient of projection matrix

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  VBPR:
    meta:
      save_recs: True
    lr: 0.0005
    epochs: 50
    factors: 100
    factors_d: 20
    batch_size: 128
    l_w: 0.000025
    l_b: 0
    l_e: 0.002

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.visual_recommenders.VBPR.VBPR_model module¶

Module description:

class elliot.recommender.visual_recommenders.VBPR.VBPR_model.VBPR_model(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None)[source]¶

get_config()[source]¶

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Returns: Python dictionary.

get_top_k(preds, train_mask, k=100)[source]¶

predict(start, stop)[source]¶

train_step(batch)[source]¶

Module contents¶

Module description:

elliot.recommender.visual_recommenders.VNPR package¶

Submodules¶

elliot.recommender.visual_recommenders.VNPR.visual_neural_personalized_ranking module¶

Module description:

class elliot.recommender.visual_recommenders.VNPR.visual_neural_personalized_ranking.VNPR(data, config, params, *args, **kwargs)[source]¶

Bases: elliot.recommender.recommender_utils_mixin.RecMixin, elliot.recommender.base_recommender_model.BaseRecommenderModel

Visual Neural Personalized Ranking for Image Recommendation

For further details, please refer to the paper

Parameters

lr – Learning rate
epochs – Number of epochs
mf_factors: – Number of latent factors for Matrix Factorization:
mlp_hidden_size – Tuple with number of units for each multi-layer perceptron layer
prob_keep_dropout – Dropout rate for multi-layer perceptron
batch_size – Batch size
l_w – Regularization coefficient

To include the recommendation model, add it to the config file adopting the following pattern:

models:
  VNPR:
    meta:
      save_recs: True
    lr: 0.001
    epochs: 50
    mf_factors: 10
    mlp_hidden_size: (32, 1)
    prob_keep_dropout: 0.2
    batch_size: 64
    l_w: 0.001

get_recommendations(k: int = 100)[source]¶

property name¶

train()[source]¶

elliot.recommender.visual_recommenders.VNPR.visual_neural_personalized_ranking_model module¶

Module description:

class elliot.recommender.visual_recommenders.VNPR.visual_neural_personalized_ranking_model.VNPRModel(*args, **kwargs)[source]¶

Bases: tensorflow.python.keras.engine.training.Model

call(inputs, training=None, mask=None)[source]¶

get_recs(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

get_top_k(preds, train_mask, k=100)[source]¶

predict(inputs, training=False, **kwargs)[source]¶

Get full predictions on the whole users/items matrix.

Returns: The matrix of predicted values.

train_step(batch)[source]¶

Module contents¶

Module description:

Submodules¶

elliot.recommender.base_recommender_model module¶

Module description:

class elliot.recommender.base_recommender_model.BaseRecommenderModel(data, config, params, *args, **kwargs)[source]¶

Bases: abc.ABC

autoset_params()[source]¶

Define Parameters as tuples: (variable_name, public_name, shortcut, default, reading_function, printing_function) Example:

self._params_list = [: (“_similarity”, “similarity”, “sim”, “cosine”, None, None), (“_user_profile_type”, “user_profile”, “up”, “tfidf”, None, None), (“_item_profile_type”, “item_profile”, “ip”, “tfidf”, None, None), (“_mlpunits”, “mlp_units”, “mlpunits”, “(1,2,3)”, lambda x: list(make_tuple(x)), lambda x: str(x).replace(“,”, “-“)),

]

abstract get_loss()[source]¶

abstract get_params()[source]¶

get_params_shortcut()[source]¶

abstract get_recommendations(*args)[source]¶

abstract get_results()[source]¶

abstract train()[source]¶

elliot.recommender.base_recommender_model.init_charger(init)[source]¶

elliot.recommender.recommender_utils_mixin module¶

class elliot.recommender.recommender_utils_mixin.RecMixin[source]¶

Bases: object

get_loss()[source]¶

get_params()[source]¶

get_recommendations(k: int = 100)[source]¶

get_results()[source]¶

get_train_mask(start, stop)[source]¶

restore_weights()[source]¶

train()[source]¶

Module contents¶

Module description:

elliot.result_handler package¶

Submodules¶

elliot.result_handler.result_handler module¶

Module description:

class elliot.result_handler.result_handler.HyperParameterStudy(rel_threshold=1)[source]¶

Bases: object

add_trials(obj)[source]¶

save_trials(output='../results/')[source]¶

save_trials_as_triplets(output='../results/')[source]¶

class elliot.result_handler.result_handler.ResultHandler(rel_threshold=1)[source]¶

Bases: object

add_oneshot_recommender(**kwargs)[source]¶

save_best_models(output='../results/')[source]¶

save_best_results(output='../results/')[source]¶

save_best_results_as_triplets(output='../results/')[source]¶

save_best_statistical_results(stat_test, output='../results/')[source]¶

class elliot.result_handler.result_handler.StatTest(value)[source]¶

Bases: enum.Enum

An enumeration.

PairedTTest = [<class 'elliot.evaluation.statistical_significance.PairedTTest'>, 'paired_ttest']¶

WilcoxonTest = [<class 'elliot.evaluation.statistical_significance.WilcoxonTest'>, 'wilcoxon_test']¶

Module contents¶

Module description:

elliot.splitter package¶

Submodules¶

elliot.splitter.base_splitter module¶

class elliot.splitter.base_splitter.Splitter(data: pandas.DataFrame, splitting_ns: types.SimpleNamespace)[source]¶

Bases: object

fold_list_generator(length, folds=5)[source]¶

generic_split_function(data: pandas.DataFrame, **kwargs) → List[Tuple[pandas.DataFrame, pandas.DataFrame]][source]¶

handle_hierarchy(data: pandas.DataFrame, valtest_splitting_ns: types.SimpleNamespace) → List[Tuple[pandas.DataFrame, pandas.DataFrame]][source]¶

process_splitting()[source]¶

read_folder(folder_path)[source]¶

rearrange_data(train_test: List[Tuple[pandas.DataFrame, pandas.DataFrame]], train_val: List[List[Tuple[pandas.DataFrame, pandas.DataFrame]]])[source]¶

splitting_best_timestamp(d: pandas.DataFrame, min_below=1, min_over=1)[source]¶

splitting_kfolds(data: pandas.DataFrame, folds=5)[source]¶

splitting_passed_timestamp(d: pandas.DataFrame, timestamp=1)[source]¶

splitting_randomsubsampling_kfolds(d: pandas.DataFrame, folds=5, ratio=0.2)[source]¶

splitting_randomsubsampling_kfolds_leavenout(d: pandas.DataFrame, folds=5, n=1)[source]¶

splitting_temporal_holdout(d: pandas.DataFrame, ratio=0.2)[source]¶

splitting_temporal_leavenout(d: pandas.DataFrame, n=1)[source]¶

store_splitting(tuple_list)[source]¶

subsampling_leavenout_list_generator(length, n=1)[source]¶

subsampling_list_generator(length, ratio=0.2)[source]¶

Module contents¶

elliot.utils package¶

Submodules¶

elliot.utils.folder module¶

Module description:

elliot.utils.folder.build_log_folder(path_log_folder)[source]¶

elliot.utils.folder.build_model_folder(path_output_rec_weight, model)[source]¶

elliot.utils.folder.create_folder_by_index(path, index)[source]¶

elliot.utils.folder.manage_directories(path_output_rec_result, path_output_rec_weight, path_output_rec_performance)[source]¶

elliot.utils.logger_util module¶

class elliot.utils.logger_util.QueueListenerHandler(handlers, respect_handler_level=False, auto_run=True, queue=<queue.Queue object>)[source]¶

Bases: logging.handlers.QueueHandler

emit(record)[source]¶

Emit a record.

Writes the LogRecord to the queue, preparing it for pickling first.

start()[source]¶

stop()[source]¶

elliot.utils.logging module¶

class elliot.utils.logging.TimeFilter(name='')[source]¶

Bases: logging.Filter

filter(record: logging.LogRecord) → bool[source]¶

Determine if the specified record is to be logged.

Is the specified record to be logged? Returns 0 for no, nonzero for yes. If deemed appropriate, the record may be modified in-place.

elliot.utils.logging.get_logger(name, log_level=10)[source]¶

elliot.utils.logging.init(path_config, folder_log, log_level=30)[source]¶

elliot.utils.logging.prepare_logger(name, path, log_level=10)[source]¶

elliot.utils.read module¶

Module description:

elliot.utils.read.find_checkpoint(dir, restore_epochs, epochs, rec, best=0)[source]¶

Parameters

dir – directory of the model where we start from the reading.
restore_epochs – epoch from which we start from.
epochs – epochs from which we restore (0 means that we have best)
rec – recommender model
best – 0 No Best - 1 Search for the Best

Returns

elliot.utils.read.load_obj(name)[source]¶: Load the pkl object by name :param name: name of file :return:

elliot.utils.read.read_config(sections_fields)[source]¶

Parameters: sections_fields (list) – list of fields to retrieve from configuration file
Returns: A list of configuration values.

elliot.utils.read.read_csv(filename)[source]¶

Parameters: filename (str) – csv file path
Returns: A pandas dataframe.

elliot.utils.read.read_imagenet_classes_txt(filename)[source]¶

Parameters: filename (str) – txt file path
Returns: A list with 1000 imagenet classes as strings.

elliot.utils.read.read_multi_config()[source]¶

It reads a config file that contains the configuration parameters for the recommendation systems.

Returns: A list of configuration settings.

elliot.utils.read.read_np(filename)[source]¶

Parameters: filename (str) – filename of numpy to load
Returns: The loaded numpy.

elliot.utils.write module¶

Module description:

elliot.utils.write.save_np(npy, filename)[source]¶: Store numpy to memory. :param npy: numpy to save :param filename: filename :type filename: str

elliot.utils.write.save_obj(obj, name)[source]¶: Store the object in a pkl file :param obj: python object to be stored :param name: file name (Not insert .pkl) :return:

elliot.utils.write.store_recommendation(recommendations, path='')[source]¶: Store recommendation list (top-k) :return:

Module contents¶

Module description:

Submodules¶

elliot.run module¶

Module description:

elliot.run.config_test(builder, base)[source]¶

elliot.run.run_experiment(config_path: str = './config/config.yml')[source]¶

Module contents¶

Module description: