The EPFL Recommender System Challenge 2019

Machine Learning - Project 2

Tshtsh_club: Marie Anselmet, Sofia Dandjee, Héloïse Monnet

Install the surprise library

pip install scikit-surprise

OR

conda install scikit-surprise

Run the project

1. Make sure that Python >= 3.7, NumPy >= 1.16 and sklearn >= 0.22 are installed
2. Go to script\ folder and run run.py. You will get submission.csv for Kaggle in the data\ folder.

cd script
python run.py

Script files

proj2_helpers.py

• calculate_mse: Computes mean-squared error between predicted and known ratings
• calculate_rmse: Computes root-mean-squared error between predicted and known ratings
• init_MF: Initializes the parameters for the custom matrix factorization methods.
• compute_error: Compute the loss (MSE) of the prediction of nonzero elements for custom matrix factorization methods.
• compute_predictions: Compute the prediction of nonzero elements for custom matrix factorization methods.
• update_user_feature: Updates user feature matrix for the ALS matrix factorization method.
• update_item_feature: Updates item feature matrix for the ALS matrix factorization method.

data_helpers.py

• read_csv_sample : Reads the sample_submission file and extracts the couples (item, user) for which the rating has to be predicted.
• create_csv_submission: Creates an output file in csv format for submission to kaggle
• build_surprise_data: Loads the training and test set for it to be usable by the surprise library
• split_data: Loads the training and test set that are used for the custom models

implementations.py

• global_mean: Use the global mean as the prediction.
• user_mean: Use the user means as the prediction.
• item_mean: Use the item means as the prediction.
• matrix_factorization_als: Matrix factorization by Alternating Least Squares (ALS).
• matrix_factorization_sgd: Matrix factorization by Stochastic Gradient Descent (SGD).
• normal_predictor: Generates predictions according to a normal distribution estimated from the training set.
• baseline_only: Combines user and item mean with user and item biases.
• knn_baseline_user: Nearest neighbour approach between users taking into account baseline ratings.
• knn_baseline_movie: Nearest neighbour approach between movies taking into account baseline ratings.
• svd: Matrix factorization algorithm taking biases into account.
• svdpp: Extension of svd taking into account implicit ratings.
• slopeone: Item-based algorithm based on similarity between users that rated the same movie.
• nmf: Non-negative matrix factorization.
• blending: Computes a ridge regression to find optimal weights for each of the fed models
• co_clustering: Users and items are assigned to clusters and co-clusters.
• knn_centered_user: Nearest neighbour approach taking into account the mean ratings of each user
• knn_centered_movie: Nearest neighbour approach taking into account the mean ratings of each movie
• knn_basic_user: Nearest neighbour basic approach on users
• knn_basic_movie: Nearest neighbour basic approach on movies

run.py

Script to produce the same .csv predictions used in the best submission on the Kaggle platform.

run_blending.py

Script which computes the predictions of the blending of our 5 best-performing models