If you want to use the original csv of the competition, run the following script to preprocess the data and save the matrices used in the various algorithms for later usage:

python preprocessing/urm.py

Matrices are saved into raw_data/matrices. If you have your own, dataset, you can skip this step.

The module Data is the main wrapper between to access the dataset saved in the disk. It is used to load the csv files and the .npz files as arrays or sparse matrices or dataframes (depending on the type of usage).

N_PLAYLISTS = 50446
N_TRACKS = 20635
N_ARTISTS = 6668
N_ALBUMS = 12744

get_urm()
get_urm_train()
get_urm_test()
get_icm()

get_tracks_df()
get_playlists_df()

get_target_playlists()
get_all_playlists()

If you want to adapt your own dataset, you have to rename your files so that they reflect the paths of this class.

All the implemented recommenders are inside the following folder: recommenders

They inherit from a common base class called RecommenderBase. This abstract class exposes some useful methods, like:

fit()

Fit the model on the data. Inherited class should extend this method in the appropriate way.

recommend(self, userid, N=10, urm=None, filter_already_liked=True, with_scores=False, items_to_exclude=[])

Compute the N best items for the specified user.

recommend_batch(self, userids, N=10, urm=None, filter_already_liked=True, with_scores=False, items_to_exclude=[], verbose=False)

Recommend the N best items for the specified list of users.

evaluate(self, recommendations, test_urm, at_k=10)

Return the MAP@k (default MAP10) evaluation for the provided recommendations computed with respect to the test_urm.

To run the models, check the various files inside: run.

Each file contains a run and a test function: test only calls run in verbose mode, without saving any result into the disk, useful to try a model and see how it behaves