Learning Fast: Home Experimentation and Open Research at Spotify
summary
Spotify personalizes the default Home for 191 million users across 40 million tracks using contextual bandits to balance exploration and exploitation, plus an internal offline framework ML Lab for reproducible evaluations and leaderboards. Home ranks cards and shelves by predicted rewards (clicks, plays, follows) and correlates offline scores with live A/B metrics. Open research released 1 million playlists for the RecSys Challenge 2018 and 130 million listening sessions for the 2019 challenge, yielding thousands of participants, hundreds of teams and over 1,000 submissions.
AI-written summary · about
chapters (10)
- 0:00Start
- 3:11Home Screen
- 6:19Presentation Bias
- 6:53Contextual Bandits
- 7:23One-Armed Bandits
- 8:47The Reward Function
- 10:14The Traditional Experimentation Process
- 11:36Offline Evaluation
- 12:02Ml Lab
- 17:48Recommended Songs
related talks
59mClosing Keynote: Paul Lamere
40mPowering Spotify's Audio Personalization Platform
04mMusic, the window into your soul
29mHow Spotify Does ML At Scale
27mPlay
50mBetter Spotify Playlists through Music Theory & Discrete Optimization
13mRevolutionizing the Music Industry with Data Science
22mPersonalizing Explainable Recommendations with Multi-objective Contextual Bandits