This project was realized as part of the competition within the Politecnico di Milano course in Recommender systems

music-recommender

Music streaming services allow users to listen to songs and to create playlists of favorite songs.

This recommender system suggests songs a user would likely add to one of her playlists based on:

other tracks in the same playlist
other playlists created by the same user
other playlists created by other users

Competition

The original unsplitted dataset includes around 1M interactions (tracks belonging to a playlist) for 57k playlists and 100k items (tracks). A subset of about 10k playlists and 32k items has been selected as test playlists and items.

The goal is to recommend a list of 5 relevant items for each playlist. MAP@5 is used for evaluation.

No restrictions on the recommender algorithm choice (e.g., collaborative-filtering, content-based, hybrid, etc.) written in any language.

Model tuning

Hold-out

make tune_with_holdout

Cross-validation

make tune_with_cv

Model evaluation

Hold-out

make holdout

Cross-validation

make cv

Recommendation

To generate the recommendations

make run