This is a project for creating music recommender.
I use an extraction from Million Song Dataset or you can try Million Song Subset which is smaller. (Contains ≈ 37% of songs from train)
MSD provides following features (aggregated in generate_data_from_msd.py)
An index file:
| song_id | artist | title |
|---|---|---|
| SOPIFJT12A81C221AC | Bob Marley & The Wailers | Three Little Birds |
And a feature-file which contains audio features that are 25 in total:
- 12 timbre features
- 12 chroma components
- max loudness value
for every segment of the track.
So every song has a numpy vector associated to it of shape (num_segmets, 25)
I use user playcounts from MSD Challenge.
- Train — Taste profile triples (user, song, count)
- Test — Visible and hidden parts from test files of EvalDataYear1. Visible is the query for user and hidden is the test. Corresponds to private part on Kaggle.
Competition used MAP@500, you can match yourself with the leaderboard if you use it.
Example:
| user | song | plays |
|---|---|---|
| 00007a02388c208ea7176479f6ae06f8224355b3 | SOAITVD12A6D4F824B | 3 |
MSD is very hard to handle because of large size (500GB) and lack of maintenance so today it is only available via AWS snapshot which requires money to download.
I don't need all the data so I got only some of the features that I needed.
I also deleted songs that were mismatched so everything is nice and clean.
You can download my version of files here:
- Million Song Subset - 1MB
- Million Song Dataset - 140MB
- Ratings - 490MB
May be someone will find this useful.