Understanding Music Playlists
- 1. Understanding Music Playlists 10 July 2015 ICML 2015 Workshop - Machine Learning and Music Discovery Keunwoo Choi George Fazekas Mark Sandler @c4dm @Queen Mary University of London
- 2. Content • Music Playlist • Finding 1 • Finding 2 • Finding 3 • Conclusion
- 3. Playlist and Recommendation • Music recommendation == playlist generation in many cases; especially for common music listener. • Because recommending a song doesn’t make sense. • Because simply picking top-N songs might fail.
- 4. Music Playlist • What is playlist? • “Sequence of music items” for ( ), by ( ), … • Ill-posed definition, inductively defined by how people use • Many people use it • 1.5B playlists on Spotify
- 5. Different Assumptions • What is a good playlist? Sequence of similar songs Smooth transitions Fixed start/end song Given duration ?
- 6. Datasets Deezer-2015 144,726 songs 50,000 playlists during 2007-2015 EchoNest track features (high-level features such as speechness, dancability, …) AoTM-2011 97,411 songs 86,310 playlists during 1998-2011 EchoNest Timbre Features, energy/ key/loudness/ mode… + playlist category
- 7. Datasets • Hierarchy of playlist categories Genre Rock Jazz Hiphop R&B Electronic Folk Rock/ Pop Mixed Genre Blues Raggae Country Punk Hardcore Dance/ House Activity Sleep Road trip Emotion Break up Depression Others Indie Alternating DJ Theme Single artist Cover Narrative
- 9. F1. User clusters in audio-based feature domain Jennings, 2007
- 13. • No structural difference • Playlist itself doesn’t represent the user that much. (Or is not easily observed.) • Usage data may be required • Usage hours, number of songs/artists, diversity of preferences, price tier, social activities, …
- 14. F2. Similarity vs. Diversity “Songs in a playlist should be similar” “Songs in a playlist shouldn’t be too similar” Similar Familiar Unified vs. Interesting Not boring Diverse
- 15. • Audio feature similarity between songs • within-playlist vs. arbitrary pairs
- 16. Mean 25% 75% 0.080 0.061 0.094 0.095 0.041 0.126
- 17. 3. Different similarity given category • Compute the similarity of songs in the playlists for each category (for each feature) • Get rankings of categories (for each feature) • Get average of the rankings • Plot it (with nice colours)
- 18. 3. Different similarity given category
- 19. Summary • User (behaviour) information is required to build user model based on clustering • Should find an appropriate range of similarity for better playlist generation • which varies given dataset, features, and similarity measure. • Desired song similarity may be different for each category • Different parameters/prior should be set.
- 20. Future work , or just curious about… • How much do people care about playlist? How much do people put efforts on it? • Mix Tape/CD was important for us (music researchers), so as (modern) playlist for people? • Looks like they are just containers for a set of songs rather than a sequence songs.