Skip to content

Latest commit

 

History

History
61 lines (47 loc) · 4.91 KB

README.md

File metadata and controls

61 lines (47 loc) · 4.91 KB
Raw

Collaborative models for Collaborative Filtering

In the age of information, users are feeling overwhelmed by countless offerings and services. Recommender systems aim to alleviate this issue by presenting personalized recommendations to the user. They are often based on Collaborative filtering (CF), which analyzes past transactions of the user in order to find new recommendations.

We propose a new model that combines the existing results of many years of research on CF. In total, we adopt four different types of approaches: RMF, Autoencoder, SVD++ and NCF. For some approaches, instead of optimizing parameters, we combine the results using an ensemble. \In addition, we introduce a probabilistic approach for Neural Collaborative filtering, which drastically outperforms current NCF models.

Authors

Nathan Corecco, Giorgio Piatti, David Gu

Overview

/AE                         -- Autoencoder
/ALS                        -- Alternating least squares (baseline)
/ALS_ensemble               -- Alternating least squares ensemble
/data                       -- Kaggle dataset
/data_val_train_kfold       -- 5 partition to compute CV score
/NCF                        -- Neural Collaborative Filtering (baseline)
/PNCF                       -- Probabilistic Neural Collaborative Filtering
/PNCF_base                  -- Probabilistic Neural Collaborative Filtering, base model (without FGE)
/results_baseline           -- Results of baseline models (need to download, see additional data section)
/results_ensemble           -- Results of sub model of the ensemble (need to download, see additional data section)
/SVDpp                      -- SVD++ (baseline)
/SVDpp_ensemble             -- SVD++ ensemble
analysis_baseline.ipynb     -- Compute score on CV split for baseline models
analysis_ensemble.ipynb     -- Compute score on CV split for the final ensemble
ensemble_lsq.ipynb          -- Compute final result of the ensemble
requirements.txt            -- Requirement file of the necessary libraries
/paper_generate_graphs      -- Code used to generate the graphs in the paper
paper.pdf                   -- paper describing our approaches

Each baseline model folder (ALS, NCF, SVDpp) shares the same general structure, namely

  • run_prepare_baseline_final.py: Computes the final prediction on the test dataset, using all available train data.
  • run_prepare_baseline_validation.py: Requires as first argument the split id(between 0 and 4). It trains the model on data_val_train_kfold/partition_k_train.csv and evaluates on the user/movie pair specified by data_val_train_kfold/partition_k_val.csv. Those 2 files are used to computed the predictions for model evaluation, which are saved for convenience under the folder results_baseline.

Each sub model folder (AE, ALS_ensemble, PNCF, SVDpp_ensemble) shares the same general structure, namely

  • run_prepare_ensemble_final.py: Computes the final prediction on the test dataset, using all available train data.
  • run_prepare_ensemble_validation.py: Requires as first argument the split id(between 0 and 4). It trains the model on data_val_train_kfold/partition_k_train.csv and evaluates on the user/movie pair specified by data_val_train_kfold/partition_k_val.csv. Those 2 files are used to computed the predictions for model evaluation and for computing the weightning coefficient for the final ensemble, which are saved for convenience under the folder results_ensemble.

Additional data

We already provided the intermediate prediciton to be able to reproduce the final ensemble result.

Submission generation

  • To reproduce the intermediate results (i.e. to compute the content of the folder results_ensemble) run the sh script prepare_ensemble.sh or on Euler euler_prepare_ensemble.sh from the main folder. Note on Euler it's best to modify the env variable BASE_DIR_RESULTS defined in the .env file to point to a scratch directory.

  • Run the notebook ensemble_lsq.ipynb, the results are stored in AE-ALS_ensemble-PNCF-SVDpp_ensemble-LSQ-results.csv

Baseline score evaluation

  • To reproduce the intermediate results (i.e. to compute the content of the folder results_baseline) run the sh script prepare_baseline.sh or on Euler euler_prepare_baseline.sh from the main folder. Note on Euler it's best to modify the env variable BASE_DIR_RESULTS defined in the .env file to point to a scratch directory.

  • Run the notebook ensemble_lsq.ipynb, the results are stored in AE-ALS_ensemble-PNCF-SVDpp_ensemble-LSQ-results.csv

Installation

To run the code, it is recommended to first run the following instructions: pip install -r requirements.txt