For the task of automatic playlist continuation described in RecSys Challenge 2018, we propose a multi-stage IR (Information Retrieval) approach combined with LTR (Learning to Rank).
Our model is capable of generating a list of recommended tracks for a given playlist which has a number of withheld tracks. Briefly, we follow three main steps to produce relevant results:
- Indexing: We build a Lucene index over the MPD.
- Searching: By treating playlists as documents and tracks as queries, we search for possible candidates over the index for an incomplete playlist.
- Ranking: We apply LambdaMART algorithm to re-rank the list of candidate tracks.
During the ranking step, we use statistical features derived from the MPD. Optionally, we can also utilize audio and album features collected via Spotify Web API. Thus, our model can compete in both main and creative tracks of the challenge. Following table presents the complete set of features:
| Feature ID | Description | Category | Track |
|---|---|---|---|
| 1 | Number of tracks in playlist | P | Main & Creative |
| 2 | Number of samples in playlist | P | Main & Creative |
| 3 | Number of holdouts in playlist | P | Main & Creative |
| 4 | Length of playlist title | P | Main & Creative |
| 5 | Track occurrence | T | Main & Creative |
| 6 | Track frequency | T | Main & Creative |
| 7 | Track duration | T | Main & Creative |
| 8 | Album occurrence | T | Main & Creative |
| 9 | Album frequency | T | Main & Creative |
| 10 | Number of tracks in album | T | Main & Creative |
| 11 | Artist occurrence | T | Main & Creative |
| 12 | Artist frequency | T | Main & Creative |
| 13 | Number of albums of artist | T | Main & Creative |
| 14 | Number of tracks of artist | T | Main & Creative |
| 15 | Submission order | PxT | Main & Creative |
| 16 | Lucene score | PxT | Main & Creative |
| 17 | Predicted position | PxT | Main & Creative |
| 18 | Jaccard distance of playlist title and track name | PxT | Main & Creative |
| 19 | Jaccard distance of playlist title and album name | PxT | Main & Creative |
| 20 | Jaccard distance of playlist title and artist name | PxT | Main & Creative |
| 21 | Audio danceability | T | Creative Only |
| 22 | Audio energy | T | Creative Only |
| 23 | Audio key | T | Creative Only |
| 24 | Audio loudness | T | Creative Only |
| 25 | Audio mode | T | Creative Only |
| 26 | Audio speechiness | T | Creative Only |
| 27 | Audio acousticness | T | Creative Only |
| 28 | Audio instrumentalness | T | Creative Only |
| 29 | Audio liveness | T | Creative Only |
| 30 | Audio valence | T | Creative Only |
| 31 | Audio tempo | T | Creative Only |
| 32 | Audio time signature | T | Creative Only |
| 33 | Album popularity | T | Creative Only |
In order to reproduce our results, one can run reproduce.sh script provided in the repository. This script builds everything needed from scratch, assuming that requirements below are available at the computing environment:
- JDK 9 or higher
- Maven 3
- Python 3.5 or later
The name of our team in the challenge is Anadolu. The members of our team are:
- Ahmet Arslan
- Alper Bilge
- Ali Yürekli