Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf ·...
Transcript of Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf ·...
![Page 1: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/1.jpg)
Music Recommendation at SpotifyHOW SPOTIFY RECOMMENDS MUSIC
Frederik Prüferxx.xx.2016
![Page 2: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/2.jpg)
• Introduction
• Explicit / Implicit Feedback
• Content-Based Filtering
• Spotify Radio
• Collaborative Filtering
• Discover Weekly Playlist
• Conclusion
2
Table of Contents
![Page 3: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/3.jpg)
Formerly
3
![Page 4: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/4.jpg)
4
Nowadays
![Page 5: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/5.jpg)
Spotify
• > 100 Mio monthly active users
• > 30 Mio songs
• > 2 billion Spotify playlists
5
![Page 6: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/6.jpg)
Spotify
„ We now have more technology than ever before to ensure that if you’re […]
doing something that only 20 people in the world will dig, we can now find those
20 people and connect the dots between the artist and listeners “ - Matthew Ogle
6
Sonnad, Nikhil. “The magic that makes Spotify’s Discover Weekly Playlist so damn good”. URL: http://qz.com/571007. [Online; last accessed June 2016]
![Page 7: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/7.jpg)
Explicit Feedback Implicit Feedback
7
• Introduction• Explicit/Implicit Feedback• Content-Based Filtering• Collaborative Filtering• Conclusion
![Page 8: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/8.jpg)
Explicit Feedback Implicit Feedback
• relies on the explicit input by
users
Actual Rating
7
![Page 9: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/9.jpg)
Explicit Feedback Implicit Feedback
• relies on the explicit input by
users
Actual Rating
7
• extracted out of the users
behavior
Predicted Rating
![Page 10: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/10.jpg)
8
Track: May 16
Artist: Lagawagon
Album: Let‘s Talk About Feelings
Release: 1998
Content-Based Filtering
![Page 11: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/11.jpg)
9
3. Content-Based Filtering
Basic Idea:
Compare sets of features which represent the items in a meaningful way
Usage:
Spotify Radio
• Introduction• Explicit/Implicit Feedback• Content-Based Filtering• Collaborative Filtering• Conclusion
![Page 12: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/12.jpg)
10
- Artist
- Album Cover
- Lyrical Meaning
- Audio Signals
- …
But which features to compare?
![Page 13: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/13.jpg)
11
Figure 1. Audio Features for a Spotify-Track
![Page 14: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/14.jpg)
The semantic gap in music
• the characteristics that affect user preference aren’t equal to the corresponding
audio signal.
• some properties are impossible to obtain from audio signals
12
![Page 15: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/15.jpg)
Predictable Recommendations
13
![Page 16: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/16.jpg)
Spotify Radio
Create a Radio Channel based on:
• Song
• Album
• Artist
• Playlist
14
![Page 17: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/17.jpg)
![Page 18: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/18.jpg)
Spotify Radio
• User plays radio -> load 250 nearest neighbors and shuffle
• Thumbs up -> Load more tracks from the thumbed-up song
• Thumbs down -> remove that song / re-weight tracks
16
![Page 19: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/19.jpg)
Nearest Neighbors
17
May 16
![Page 20: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/20.jpg)
Spotify Radio
• User plays radio -> load 250 nearest neighbors and shuffle
• Thumbs up -> Load more tracks from the thumbed-up song
• Thumbs down -> remove that song / re-weight tracks
18
![Page 21: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/21.jpg)
Collaborative Filtering
19
Basic Idea:
What do other people (with a similar music taste) listen to?
Usage:
Discovery Weekly Playlist
• Introduction• Explicit/Implicit Feedback• Content-Based Filtering• Collaborative Filtering• Conclusion
![Page 22: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/22.jpg)
20
B:A:
![Page 23: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/23.jpg)
A very big matrix…
21
30mio items
100mio users
![Page 24: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/24.jpg)
Nearest Neighbors
22
Person A
![Page 25: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/25.jpg)
Implicit Matrix Factorization
23
Users
Songs
min𝑥,𝑦
𝑢,𝑖
𝑐𝑢𝑖 𝑝𝑢𝑖 − 𝑥𝑢𝑇 ∗ 𝑦𝑖
2 + λ(
𝑢
𝑥𝑢2 +
𝑖
𝑦𝑖2)
• 𝑝𝑢𝑖 =1, 𝑢𝑠𝑒𝑟 𝑢 ℎ𝑎𝑠 𝑙𝑖𝑠𝑡𝑒𝑛𝑒𝑑 𝑡𝑜 𝑠𝑜𝑛𝑔 𝑖
0, 𝑢𝑠𝑒𝑟 𝑢 ℎ𝑎𝑠𝑛′𝑡 𝑙𝑖𝑠𝑡𝑒𝑛𝑒𝑑 𝑡𝑜 𝑠𝑜𝑛𝑔 𝑖
• Goal : 𝑥𝑢*𝑦𝑖 = 𝑝𝑢𝑖
• 𝑐𝑢𝑖 : describes the confidence, that user u likessong i
• λ : regularization penalty to avoid overfitting
≈ X Yቐ
ቐ
f
e
![Page 26: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/26.jpg)
Alternating Least Squares
1. Initialize user & item vectors
2. Fix item vectors and solve for optimal user vectors
- Take the derivative of loss function with respect to user‘s vector,
set equl to 0 and solve
3. Fix user vectors and solve for optimal item vectors
4. Repeat until convergence
24
![Page 27: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/27.jpg)
25
B:A:
![Page 28: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/28.jpg)
26
The Harry-Potter Effect:
![Page 29: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/29.jpg)
26
The Cold-Start Problem:The Harry-Potter Effect:
![Page 30: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/30.jpg)
![Page 31: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/31.jpg)
Discover Weekly Playlist
28
![Page 32: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/32.jpg)
Discover Weekly Playlist
29
Spotifyfinds songs
that fit your
profile
DiscoverWeekly Playlist
- ….- ….- ….- ….- ….- ….- ….
Spotifyidentifies
similarsongs
Develops„Taste
Profile“
You listen toand save songs
Other users createbillions of Playlists
![Page 33: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/33.jpg)
30
Figure 2. The Spotify Blob
![Page 34: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/34.jpg)
Conclusion
Music Profile Content
Other Songs Content
31
Content-Based Filtering
• Introduction• Explicit/Implicit Feedback• Content-Based Filtering• Collaborative Filtering• Conclusion
![Page 35: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/35.jpg)
Conclusion
Music Profile Content Music Profile Other users
Other Songs Content Songs they listened to
31
Content-Based Filtering Collaborative Filtering
![Page 36: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/36.jpg)
Conclusion
32
![Page 37: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/37.jpg)
References
• Balabanović, Marko, and Yoav Shoham. "Fab: content-based, collaborative recommendation." Communications of the ACM 40.3 (1997): 66-72.
• Bernhardsson, Erik. “Recommendations at Spotify v4” [Presentation Slides], URL: http://de.slideshare.net/MrChrisJohnson/algorithmic-music-recommendations-at-spotify, [Online; last accessed June 2016]
• Bernhardsson, Erik. “Nearest neighbor methods and vector models – part 1“. URL: https://erikbern.com/2015/09/24/nearest-neighbor-methods-vector-models-part-1/. [Online; last accessed June 2016]
• Dieleman, Sander. “Recommendig music on Spotify with deep learning”. URL: http://benanne.github.io/2014/08/05/spotify-cnns.html August 05, 2014 . [Online; last accessed June 2016].
• 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.
• Koren, Yehuda, Robert Bell, and Chris Volinsky. "Matrix factorization techniques for recommender systems." Computer 8 (2009): 30-37.
33
![Page 38: Music Recommendation at Spotifyhpac.cs.umu.se/teaching/sem-mus-16/presentations/Pruefer.pdf · Spotify „ We now have more technology than ever before to ensure that if youre […]](https://reader033.fdocuments.net/reader033/viewer/2022060605/605a113b0c999752382adcea/html5/thumbnails/38.jpg)
References
• Pazzani, Michael J., and Daniel Billsus. "Content-based recommendation systems." The adaptive web. Springer Berlin Heidelberg, 2007. 325-341.
• Spotify Developers. Audio Features [Table File]. URL: https://developer.spotify.com/web-api/console/get-audio-features-track/#complete. [Online; last accessed June 2016]
• Spotify Webpage. N.p. URL: https://www.spotify.com/de/. [Online; last accessed June 2016]
• Statista. N.p. URL: http://www.statista.com/statistics/244995/number-of-paying-spotify-subscribers/. [Online; last accessed June 2016]
• Steck, Harald, van Zwol, Roelof, Johnson, Chris. “Interactive Recommender Systems” [Presentation Slides], URL: http://de.slideshare.net/MrChrisJohnson/interactive-recommender-systems-with-netflix-and-Spotify, [Online; last accessed June 2016]
• Sonnad, Nikhil. The Spotify Blob [Image File]. URL: http://qz.com/571007. [Online; last accessed June 2016]
• Van den Oord, Aaron, Sander Dieleman, and Benjamin Schrauwen. "Deep content-based music recommendation." Advances in Neural Information Processing Systems. 2013.
33