Should We Care About Content? Recommending by Proxy with Big Metadata

When constructing a music recommender system, which is more important: a musicological understanding of the catalog of music in a system or the number of times two particular songs were played one after the other and were `liked’? Even better, if a system knows the latter, does the former even matter? Do machines that predict behavior need to learn to listen? Or is observing behavior enough?

Transcript

1. Should we care
   about content?
   Recommending by proxy with big metadata
   B e n F i e l d s
   @ a l s o t h i n g s
   

   View Slide

2. Goal:
   

   View Slide

3. how to make a
   recommender
   

   View Slide

4. not so much how
   

   View Slide

5. Rather: ‘what’
   

   View Slide

6. what data should be used
   to make a recommender?
   

   View Slide

7. First, Examples
   

   View Slide

8. Image
   

   View Slide

9. http://www.flickr.com/photos/dexxus/5652914929/
   

   View Slide

10. http://www.flickr.com/photos/msojka/5285298402/
    

    View Slide

11. http://www.flickr.com/photos/pinkertons/7730547912/
    

    View Slide

12. http://www.flickr.com/photos/nicholas_t/626009491/
    

    View Slide

13. http://www.flickr.com/photos/dexxus/5652914929/
    

    View Slide

14. http://www.flickr.com/photos/dexxus/5652914929/ http://www.flickr.com/photos/msojka/5285298402/
    http://www.flickr.com/photos/pinkertons/7730547912/ http://www.flickr.com/photos/nicholas_t/626009491/
    

    View Slide

15. http://www.flickr.com/photos/dexxus/5652914929/ http://www.flickr.com/photos/msojka/5285298402/
    http://www.flickr.com/photos/pinkertons/7730547912/ http://www.flickr.com/photos/nicholas_t/626009491/
    Same Camera Model
    Colour Palette Distance
    Location
    

    View Slide

16. Music
    

    View Slide

17. A small playlist
    

    View Slide

18. What should follow
    this?
    Meat Loaf -
    I'd Do Anything for Love
    

    View Slide

19. maybe?
    Beethoveen’s Piano Sonata No. 14
    in C Sharp Minor, Op. 27, No. 2
    

    View Slide

20. how about?
    Journey - Don’t Stop Believin’
    

    View Slide

21. A small playlist
    Meat Loaf -
    I'd Do Anything for Love
    Beethoveen Journey
    

    View Slide

22. Recommending?
    

    View Slide

23. prediction of opinions
    

    View Slide

24. prediction of opinions
    about things
    

    View Slide

25. prediction of Actions
    on things
    

    View Slide

26. Metadata?
    

    View Slide

27. Data About Data
    

    View Slide

28. Content descriptors
    are metadata
    

    View Slide

29. Ratings are metadata
    

    View Slide

30. But it isn’t always
    neat and Tidy
    

    View Slide

31. human
    curated
    Content descriptors
    are metadata
    

    View Slide

32. In the wild
    

    View Slide

33. @alsothings 33
    

    View Slide

34. @alsothings 34
    

    View Slide

35. @alsothings 35
    

    View Slide

36. @alsothings 36
    

    View Slide

37. @alsothings 37
    

    View Slide

38. @alsothings 38
    

    View Slide

39. View Slide

40. Contests
    

    View Slide

41. Netflix Prize
    

    View Slide

42. Netflix Prize
    500K people’s ratings
    of 18k movies
    Take these and predict
    the ratings of movies
    that haven’t been
    Rated
    

    View Slide

43. BellKor
    

    View Slide

44. gradient boosted
    decision trees
    

    View Slide

45. Any Value in
    Content?
    

    View Slide

46. No.
    

    View Slide

47. Million Song
    Dataset (kaggle)
    

    View Slide

48. Million Song Dataset
    challenge (kaggle)
    The partial listening
    history of 1M people
    Predict the tracks
    that are missing
    

    View Slide

49. Collaborative
    filtering
    

    View Slide

50. Any Value in
    Content?
    

    View Slide

51. No.
    

    View Slide

52. so?
    

    View Slide

53. moar data >
    deeper Analysis
    

    View Slide

54. Simple use of
    metadata from far
    and wide gets you
    further than Deep
    understanding of
    core-data
    

    View Slide

55. But it doesn’t
    help you explain
    things to the
    user.
    

    View Slide

56. http://www.netflixprize.com/community/viewtopic.php?id=1537
    http://www.kaggle.com/c/msdchallenge
    http://mir-in-action.blogspot.co.uk/2012/09/collaborative-
    filtering-still-rules.html
    http://www.dtic.upf.edu/~ocelma/PhD/
    http://www.slideshare.net/plamere/music-recommendation-
    and-discovery
    Ben Fields
    @alsothings
    Questions?
    

    View Slide