Machine Learning and Big Data for Music Discovery at Spotify
17 likes20,945 views
The document discusses how Spotify utilizes machine learning and big data to enhance music discovery for its 100 million users across 60 markets. It highlights various recommendation approaches including editorial, algorithmic, and collaborative filtering, as well as challenges like cold-start issues for new users and tracks. Additionally, it emphasizes the use of deep learning techniques such as convolutional neural networks to analyze audio data and generate recommendations.
Machine Learning and Big Data for Music Discovery at Spotify
- 1. Machine Learning & Big Data for Music Discovery Galvanize NYC, Mar 9th, 2017 Vidhya Murali @vid052 Ching-Wei Chen @cweichen
- 2. 100M users in 60 markets 50M subscribers Over 30M songs, and 2B playlists $5B paid to rightsholders Spotify Music for everyone.
- 6. Discover Discover Weekly How What to recommend?
- 7. Many flavors of recommendations Radio
- 8. Many flavors of recommendations Daily MixRadio
- 9. Many flavors of recommendations This Is:Daily MixRadio
- 10. Many flavors of recommendations Recommended SongsThis Is:Daily MixRadio
- 12. ‣ Editorial ‣ Algorithmic ○Content-based ■ Metadata ■ Audio Signals ○Collaborative ■ Usage based Recommendation approaches
- 13. ‣ Editorial ‣ Algorithmic ○Content-based ■ Metadata ■ Audio Signals ○Collaborative ■ Usage based ‣ Algotorial Recommendation approaches
- 14. ‣ Editorial ‣ Algorithmic ○Content-based ■ Metadata ■ Audio Signals ○Collaborative ■ Usage based ‣ Algotorial Recommendation approaches
- 15. ‣ Find patterns from user’s past behavior to generate recommendations. ‣ Domain independent ‣ Scalable Collaborative Filtering
- 16. Compact representation for each user and item (song): f-dimensional vectors Latent Factor Models
- 17. NLP Models
- 18. Context & Co-occurrence is key!
- 19. Document : Playlist Word : Song NLP Models work great on playlists!
- 21. Music in Latent Space
- 23. Music + Math = Epic
- 24. Songs as vectors
- 25. Recommendations User Profile: ● Aggregation over user interactions on Spotify ● Clustering to capture distinct user tastes/ contexts ● Time Sensitive profiling
- 26. ‣ Scale of catalog ● 30M tracks; 2B playlists ● Training ○ 25B data points ○ 100M users ○ 60 countries represented Challenges unique to spotify
- 27. Data Pipelines
- 28. Data Pipelines Big Table Big Table GCS D A T A F L O W Pub Sub Scio
- 29. ‣ Scale of catalog ● 30M tracks; 2B playlists ● Training ○ 25B data points ○ 100M users ○ 60 countries represented ‣ Cold-Start ○ New Users ○ New Music Challenges unique to spotify
- 31. What’s in a sound?
- 32. What’s in a sound? AmplitudeTime Frequencies Loudness
- 33. What’s in a sound? Melody Beats Chords Voices Instruments Lyrics AmplitudeTime Frequencies Loudness
- 34. What’s in a sound? Melody Beats Chords Voices Instruments Lyrics AmplitudeTime Frequencies Loudness Popularity Era Region Genre Mood Purpose
- 36. Deep Learning
- 37. Deep Learning 1. No feature extraction necessary 2. LOTS of simple learning nodes in many layers 3. Propogate errors backwards to learn optimal weights 4. Needs LOTS of data
- 39. Typical Convolutional Neural Network
- 40. Deep Learning on Audio at Spotify Sander Dieleman: http://benanne.github.io/2014/08/05/spotify-cnns.html Input: Audio spectrogram Output: Latent Space Vector
- 42. Cold Start? Problem solved! * * Not completely, of course!
- 43. Recommending new music Release Radar Fresh Finds
- 44. Recommendations at Spotify Recommended SongsThis Is:Daily MixRadio Discover Weekly Release Radar
- 45. What’s next? ?
- 46. Join the band! www.spotify.com/jobs Ching-Wei (@cweichen): cw@spotify.com Vidhya (@vid052): vidhya@spotify.com