Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

눈으로 듣는 음악 추천 시스템

590 views

Published on

최규민(pi.314) / kakao corp.(추천팀)
---
발표의 시작은? Music Streaming 서비스인 멜론과 카카오미니에 적용된
Toros 음악 추천 시스템의 다양한 추천 모델과 이를 활용한 추천 레시피에 대한 이야기를 합니다.

그리고 마무리는? 음악 추천을 위해 40~100차원 벡터로 모델링 된 (CB/CF) Latent Feature들을
다양한 조합으로 Visualization(t-SNE)하고 눈으로 탐색해 가면서 재미난 특징을 찾아보고자 합니다.
"CB와 CF Feature 노래를 어떻게 표현했을까?”,
"Popular노래와 Rare 한 노래의 Feature는?”
“Feedback이 많은 유저와 적은 유저 차이는?”
“팬덤이 있는 유저들은 어떤 모양일까?"

Published in: Software
  • Be the first to comment

눈으로 듣는 음악 추천 시스템

  1. 1. ~
  2. 2. ~
  3. 3. ~
  4. 4. ~ Matrix Factorization Word2Vec Deep Learning on audio
  5. 5. ~
  6. 6. .. Dance the night 1 Power up Forever young Viva la viva  Thinking out Loud  Counting Stars Something Just Like This
  7. 7. .. Dance the night 1 Power up Forever young Viva la viva  Thinking out Loud  Counting Stars Something Just Like This
  8. 8. .. Dance the night 1 Power up Forever young Viva la viva  Thinking out Loud  Counting Stars Something Just Like This
  9. 9. Push 1.5 5.6 User 18 DJ 3 / Song Vector .
  10. 10. 1.5 5.6 User 18 DJ 3 / Song Vector .
  11. 11. 1.5 5.6 User 18 DJ 3 / Song Vector . Feedback CF(Collaboration Filter) Contents CBF(Contents Based Filtering)
  12. 12. ?
  13. 13. ? …
  14. 14. ? filtering . … / 7080 X- 2000’ 2010’ 2018’ Top100 Rare - …..
  15. 15. 40D Dense Vector - Matrix Completion via ALS - User/Item Latent Feature Feedback Log(8day) - 4.3M(User) X 2.1M(Song) - 388M None zero values (Song)
  16. 16. 40D Dense Vector - Matrix Completion via ALS - User/Item Latent Feature Feedback Log(8day) - 4.3M(User) X 2.1M(Song) - 388M None zero values (Song) (Song) Vector Vector .
  17. 17. ? , .
  18. 18. Low level Feature Genre Classification Network
  19. 19. Low level Feature Genre Classification Network(Song) Vector Vector .
  20. 20. - 2004 12 1553
  21. 21. - 2004 12 1553 Stream + W2V
  22. 22. - 2004 12 1553 2004 12 Stream + W2V
  23. 23. - 2004 12 1553 Stream + MF 1553
  24. 24. Sampling 5M -> 16K Song Meta Mapping , , Stream Song Segmentation Light, Medium, Heavy, Extreme Dimension Reduction 40D -> 2D
  25. 25. Sampling 5M -> 16K Song Meta Mapping , , Stream Song Segmentation Light, Medium, Heavy, Extreme Dimension Reduction 40D -> 2D 5M Songs Top 1,000 Songs 15,000 Songs Uniform Sampling 1,000 Songs + = 16,000 Songs
  26. 26. Sampling 5M -> 16K Song Meta Mapping , , Stream Song Segmentation Light, Medium, Heavy, Extreme Dimension Reduction 40D -> 2D t-SNE
  27. 27. Sampling 5M -> 16K Song Meta Mapping , , Popularity Song Segmentation Light, Medium, Heavy, Extreme Dimension Reduction 40D -> 2D
  28. 28. Sampling 5M -> 16K Song Meta Mapping , , Popularity Song Segmentation Light, Medium, Heavy, Extreme Dimension Reduction 40D -> 2D
  29. 29. • Heavy • Top1000 Top1000 . • Popularity . • .
  30. 30. OST R&B CCM J-POP POP Global
  31. 31. Popularity
  32. 32. Light, Medium Song
  33. 33. Top1000 Top1000
  34. 34. • • MF • Top1000 Top1000 .
  35. 35. • • • Top1000 Top1000 .
  36. 36. 1,037 .
  37. 37. 45.2% 60.1% 59.9%
  38. 38. Stream Stream Stream
  39. 39. Stream Stream Stream 1. Streaming 1 2. Streaming 1 Streaming 97
  40. 40. Stream 1. 100 ~ 10,000 . 2. .
  41. 41. Song Vector Popular .
  42. 42. / / CCM /
  43. 43. .

×