Presentation slide for the paper "Towards Seed-Free Music Playlist Generation: Enhancing Collaborative Filtering with Playlist Title Information", which is accepted as presenting team for the RecSys Challenge 2018.

1. Towards Seed-Free Music Playlist
Generation:
Enhancing Collaborative Filtering with Playlist Title
Information
Jaehun Kim, Minz Won, Cynthia C. S. Liem, Alan Hanjalic
1

2. Preliminary Approach

3. First attempt : WRMF
3
● Good Old MF
○ Weighted Regularized Matrix Factorization [1]
■ Developed for implicit feedback
■ ALS* optimization : fast and reliable
■ Only 2~3 hyper parameters
*Alternating Least Square (or Coordinate Descent)
R U~= x V

4. First attempt : WRMF
4

5. First attempt : WRMF
5

6. First attempt : WRMF
6
NO SEED
Average
Winner

7. First attempt : WRMF
7
● Good Old MF
○ Already reasonable performance
○ Except No-Seed case => Cold Start Problem
● Any metadata or content for playlist?

8. First attempt : WRMF
8
● Good Old MF
○ Already reasonable performance
○ Except No-Seed case => Cold Start Problem
● Any metadata or content for playlist?
○ Playlist titles!

9. Playlist Titles

10. ● Text information (implicitly) represents the playlist
● Some key statistics
Playlist Titles
10
# of titles (MPD + Challenge Set) 1,010,000
# of unique titles 93,250
# of unique titles (stemmed) 49,808
Single word ~60%
Less than two words ~92%

11. ● Text information (implicitly) represents the playlist
● Some key statistics
Playlist Titles
11
# of titles (MPD + Challenge Set) 1,010,000
# of unique titles 93,250
# of unique titles (stemmed) 49,808
Single word ~60%
Less than two words ~92%
1. Playlist titles ~= Words

12. ● Noisiness
Playlist Titles
12
Categories Examples
Special characters //Pretty Little Liars//, ** some tunes, ?!?
Repeated characters Yaaaas, summerrrr, partayyy
Shortened words Chillin, Temp, favss
Abbreviated words loml, jb, IDFK, jjjj
Symbolic expressions
Multiple languages otoño, 電台收藏, アニメ

13. ● Noisiness
Playlist Titles
13
Categories Examples
Special characters //Pretty Little Liars//, ** some tunes, ?!?
Repeated characters Yaaaas, summerrrr, partayyy
Shortened words Chillin, Temp, favss
Abbreviated words loml, jb, IDFK, jjjj
Symbolic expressions
Multiple languages otoño, 電台收藏, アニメ
2. Standard word-level approaches
(may be) NOT WORKING

14. Playlist Titles
14
● Playlist titles ~= words
● Standard word-level approaches (may be) not working
● Character level approach : Character N-GRAM

15. Character N-gram
15
Input Text “Character N-gram”
Unigram (1-gram) C, h, a, r, a, c, t, e, r, , N, -, g, r, a, m
Bigram (2-gram) Ch, ha, ar, ac, ct, te, er, r , N, N-, -g, ...
Trigram (3-gram) Cha, har, arc, rct, cte, ter, er , r N, N-, ...
Quadrogram (4-gram) Char, harc, arct, rcte, cter, ter , er N, ...
... ...

16. Character N-gram
16
Input Text “Character N-gram”
Unigram (1-gram) C, h, a, r, a, c, t, e, r, , N, -, g, r, a, m
Bigram (2-gram) Ch, ha, ar, ac, ct, te, er, r , N, N-, -g, ...
Trigram (3-gram) Cha, har, arc, rct, cte, ter, er , r N, N-, ...
Quadrogram (4-gram) Char, harc, arct, rcte, cter, ter , er N, ...
... ...
Cha har arc cte ter er r N N-abc zeb ytj pyk nrc nfe
1 1 0 1 0 1 0 1 0 1 0 1 0 1 0 ...
jkl
Bag-of-Character-N-gram

17. ● Build bag-of-n-grams for each playlist (Train + Test Set)
● For each testing playlist
○ Find M closest playlist in Train set using cosine distance
○ Collect tracks from retrieved playlist
○ Recommend L most popular tracks
Title-based RecSys
NGRAM:Similarity Based
17

18. Cosine
Distance
Title-based RecSys
NGRAM:Similarity Based
18
Cha har arc cte ter er r N N-abc zeb ytj pyk nrc nfe
1 1 0 1 0 1 0 1 0 1 0 1 0 1 0 ...
jkl
...
...
Testing
Playlists
Training
Playlist
L most popular tracks among
M closest training playlists
Bag-of-Character-N-gram

19. Title-based RecSys
:: Model Based
19

20. Title-based RecSys
:: Model Based
20
1.
(pre-trained)
Transfer`Title info’

21. Title-based RecSys
:: Model Based
21
1. 2.
Substitute

22. Title-based RecSys
:: Model Based
22
1. 2.
?X. Wang et al. (2014) [2] A. Van den Oord et al. (2013) [3]

23. Title-based RecSys
:: Model Based
23
1. 2.
+
?!

24. Proposed Model

25. Multi-Objective Collab. Filtering
25
MF for Pre-trained Factors
Main objective function for

26. Multi-Objective Collab. Filtering
26
1. 2.

27. Multi-Objective Collab. Filtering
27
1. 2.

28. RNCF : Recurrent Neural CF
28

29. Results

30. Results:Overall
30
● It works! (10th on final Leaderboard)

31. Results:Overall
31
● Superior over simple baselines

32. Results:Overall
32
● WRMF > SVD

33. HRNCF : Hybrid RNCF
33
R U~= x V
Filling missing
factors

34. Results:Overall
34
● HRNCF: best of both worlds

35. Results:Overall
35
● Simple NGRAM distance may have
given us very similar performance

36. Hyperparameters: WRMF
36

37. Hyperparameters: WRMF
37

38. Hyperparameters: RNCF
38

39. Hyperparameters: RNCF
39

40. Hyperparameters: RNCF
40

41. Discussion & Take away

42. Take away
42
● MF is still powerful
● Setting up right (internal) evaluation setup is more important than model
● Software engineering DOES MATTER
○ Since the scalability DOES MATTER
○ Since hyper-parameter tuning DOES MATTER
● Deep learning is not a magic wand
○ No Free Lunch
○ It costs a LOT
● Content-based algorithms still gives small (but significant) gain to CF

43. Thank you!
Code: https://github.com/eldrin/recsys18-spotify-spotif-ai
43

44. References
44
[1] Hu, Yifan, Yehuda Koren, and Chris Volinsky. "Collaborative filtering for implicit feedback datasets." Data Mining, 2008.
ICDM'08. Eighth IEEE International Conference on. Ieee, 2008.
[2] Wang, Xinxi, and Ye Wang. "Improving content-based and hybrid music recommendation using deep learning."
Proceedings of the 22nd ACM international conference on Multimedia. ACM, 2014.
[3] Van den Oord, Aäron, Sander Dieleman, and Benjamin Schrauwen. "Deep content-based music recommendation."
Advances in neural information processing systems. 2013.