Music r ecommendations at Spotify

1. Music recommendations at Spotify Erik Bernhardssonerikbern@spotify.com

2. Spotify • Launched in 2009 • Available in 17 countries • 20M active users, 5M paying subscribers • Peak at 5k tracks/s, 1M logged in users • 20M tracks

3. Some applications

4. Recommendation stuff at Spotify • Related artists:

5. Recommendation stuff at Spotify, cont…

6. More!

7. How can we find music?

8. Recommendations • Manual classification • Feature extraction • Social media analysis, web scraping, metadata based • Collaborative filtering

9. Pandora & Music Genome Project • Classifies tracks in terms of 400 attributes • Each track takes 20-30 minutes to classify • A distance function finds similar tracks • “Subtle use of strings” • “Epic buildup” • “Acid Jazz roots” • “Beats made for dancing” • “Trippy soundscapes” • “Great trombone solo” • …

10. Scraping the web is another approach

11. Feature extraction

12. Collaborative filtering • Idea: • If two movies x, y get similar ratings then they are probably similar • If a lot of users all listen to tracks x, y, z, then those tracks are probably similar

13. Collaborative filtering

14. Get data

15. … lots of data

16. Aggregate data • Throw away temporal information and just look at the number of times

17. OK, so now we have a big matrix

18. … very big matrix • Throw out all the temporal data:

19. Supervised collaborative filtering is pretty much matrix completion

20. Supervised learning: Matrix completion

21. Supervised: evaluating rec quality

22. Unsupervised learning • Trying to estimate the density • i.e. predict probability of future events

23. Try to predict the future given the past

24. How can we find similar items

25. We can calculate correlation coefficient as an item similarity • Use something like Pearson, Jaccard, …

26. Amazon did this for “customers who bought this also bought” • US patent 7113917

27. Parallelization is hard though

28. Can speed this up using various LSH tricks • Twitter: Dimension Independent Similarity Computation (DISCO)

29. Are there other approaches?

30. Natural Language Processing has a lot of similar problems • …matrix factorization is one idea

31. Matrix factorization

32. Matrix factorization • Want to get user vectors and item vectors • Assume f latent factors (dimensions) for each user/item

33. Probabilistic Latent Semantic Analysis (PLSA) • Hofmann, 1999 • Also called PLSI

34. PLSA, cont. • + a bunch of constraints:

35. PLSA, cont. • Optimization problem: maximize log-likelihood

36. PLSA, cont.

41. “Collaborative Filtering for Implicit Feedback Datasets” • Hu, Koren, Volinsky (2008)

42. “Collaborative Filtering for Implicit Feedback Datasets”, cont.

43. Here is another method we use

44. What happens each iteration • Assign all latent vectors small random values • Perform gradient ascent to optimize log-likelihood

45. Calculate derivative and do gradient ascent • Assign all latent vectors small random values • Perform gradient ascent to optimize log-likelihood

46. 2D iteration example

47. Vectors are pretty nice because things are now super fast • User-item score is a dot product: • Item-item similarity score is a cosine similarity: • Both cases have trivial complexity in the number of factors f:

48. Example: item similarity as a cosine of vectors

49. Two dimensional example for tracks

50. We can rank all tracks by the user’s vector