Download as PDF, PPTX
Uploaded byEsh Vckay
Music Personalization : Real time Platforms.
1. The document discusses music personalization techniques at Spotify, including understanding users and music content, using collaborative filtering and latent vector models to make recommendations, and building real-time recommendation systems using Apache Storm. 2. It describes how Spotify uses machine learning techniques like matrix factorization and word2vec to generate latent vectors for users, songs, artists and playlists to measure similarity and make personalized recommendations at scale for its 75 million users. 3. The key challenges are processing huge amounts of data from 1 billion playlists and 1TB of logs daily to provide recommendations for each new user within 3 seconds and in real-time as listening behaviors change.
More Related Content
Similar to Music Personalization : Real time Platforms.
![[DSC Europe 25] Dusan Jovicic - AI Story: From on-prem to cloud and back agai...](https://cdn.slidesharecdn.com/ss_thumbnails/8kp49m6uq22ifnbwhfnk-2-251205085715-964d11a6-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Milan Sekuloski - Data, Defence, and Development: Cybersecuri...](https://cdn.slidesharecdn.com/ss_thumbnails/dfrkwwx4qly6atqpbl4z-4-251209104645-c3d4b0ca-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Jovan Bogicevic - Legacy to AI-Driven Defense: Transforming D...](https://cdn.slidesharecdn.com/ss_thumbnails/rsarluadt563hntyfc8q-3-251211083849-3e7bc4c0-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Goran Obradovic - The Rise of Sovereign AI: Building the Regi...](https://cdn.slidesharecdn.com/ss_thumbnails/7nw2xxixrxqdxvrb5wca-6-251205085714-ab09a2ac-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Dusan Nesic - Securing Tomorrow’s Infrastructure: Why Cyber-P...](https://cdn.slidesharecdn.com/ss_thumbnails/qikbszfftyowjm2q6duw-1-251211083848-8f2ead6b-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Vladimir Jelic - The AI-Driven Security Shift From Reactive D...](https://cdn.slidesharecdn.com/ss_thumbnails/6g5gj25mtjwayniqem1t-6-251209104645-7a5a5fc6-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Sara Polak - The Archaeology of Innovation: AI as the Next Cr...](https://cdn.slidesharecdn.com/ss_thumbnails/7ecbscdnt8mlcuqbd2ln-2-sara-polak-ai-creative-industries-251208152533-aa1fcf54-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Sara Polak - The Ancient Operating System: What Archaeology T...](https://cdn.slidesharecdn.com/ss_thumbnails/3vch2p6tttdnwhsgazoz-3-sara-polak-smart-cities-251208152532-64404202-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Branko Dzakula - From Defense to Attack: How AI Redefines Cyb...](https://cdn.slidesharecdn.com/ss_thumbnails/80bdzdxpr3ky2g0qvyk9-8-251211083048-ce5fc1ee-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Marija Vlajkovic & Andrea Radonjanin - Integration of AI tool...](https://cdn.slidesharecdn.com/ss_thumbnails/qf1jrglttoc3bm8s3aop-final-integration-of-ai-tools-251208151905-394f3a6a-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Dusan Pavlov - There Is No Spoon: Inferring Vision from Neura...](https://cdn.slidesharecdn.com/ss_thumbnails/wg0v1umoqjm4nnbd3p0v-there-is-no-spoon-251205085715-6d81d6c5-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Imai Jen-La Plante - The New Generation: AI and the Future of...](https://cdn.slidesharecdn.com/ss_thumbnails/kxi8t2l5rggivgcenyba-1-jenlaplante-dsc-251208152532-d1e076c2-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 25] Katherine Forrest - AI NOW: Understanding the Velocity of Cha...](https://cdn.slidesharecdn.com/ss_thumbnails/wvvbruqfrci0sfq9xwgb-4-251212104007-e5ad1987-thumbnail.jpg?width=640&height=640&fit=bounds)
Music Personalization : Real time Platforms.
- 1. Music Personalization: Realtime Platforms ♫+ ML + You = ❤ CrunchConf, Budapest, October 30, 2015
- 2. Esh Kumar Machine Learning& Data Products @ Spotify NYC @eshvk
- 3. Who am I? •UT Austin Machine Learning • Building Large Scale Recommendation Systems @ Mozilla, StumbleUpon & Spotify
- 4. 75 M+ ActiveUsers
- 5.
- 6. 1 TB ofLogs/Day
- 7. 1200+ Node HadoopCluster
- 8. Products •Discover … tofind new albums •Discover Weekly … A weekly Playlist •Editorial Playlist Recommendations •Radio
- 9. Music Personalization •Understanding People ➡User Experience, Cultural Variations •Understanding Content ➡ Genres, Cultural knowledge •Models ➡ Collaborative Filtering, Content Based ML Content User
- 10. Music Personalization •Understanding People ➡User Experience, Cultural Variations •Understanding Content ➡ Genres, Cultural knowledge •Models ➡ Collaborative Filtering, Content Based • News, Blogs, NLP
- 11. Music Personalization •Understanding People ➡User Experience, Cultural Variations •Understanding Content ➡ Genres, Cultural knowledge •Models ➡ Collaborative Filtering, Content Based • News, Blogs, NLP • Manually tag attributes • Curation
- 12. Music Personalization •Understanding People ➡User Experience, Cultural Variations •Understanding Content ➡ Genres, Cultural knowledge •Models ➡ Collaborative Filtering, Content Based • News, Blogs, NLP • Manually tag attributes • Curation • CF
- 13. 30 Million Songs… WhatToPlay? 75 Million Users … 1 Person Every 3 Secs…
- 14. Recommendation Systems • Predictuser response to options. • Rich field: Matrix completion, ranking, text models, latent factor models. • Several conferences annually. RecSys, NIPS, ICML etc • Industry researchers include NFLX, GOOG, MS and more…
- 15. Collaborative Filtering Hey, I liketracks P, Q, R, S! Well, I like tracks Q, R, S, T! Then you should check out track P! Nice! Btw try track T! Model you based on songs you played… Predict your future based on similar users… Millions of users and billions of streams… …. so there is someone like you out there
- 16. Collaborative Filtering The NetflixPrize. A million dollars for beating NFLX’s best algorithms by ~ 10%.
- 17. Similarity Our problem isto figure out how similar two items are. Mathematically, this means modeling a function Similarity(x,y) for all users and items, if possible.
- 18. How do wedo this? Matrix Completion. A matrix expresses a system. We model the data in the form of a matrix. For example, play counts for all songs and all users could be: Users 8 >>>>>>< >>>>>>: 0 B B B B B B @ Song Plays z }| { s1,1 s1,2 14 · · · s1,n s2,1 s2,2 2 · · · s2,n · · · sm,1 sm,2 1 · · · sm,n 1 C C C C C C A Users 8 >>>>>>< >>>>>>: 0 B B B B B B @ Song Plays z }| { s1,1 s1,2 14 · · · s1,n s2,1 s2,2 2 · · · s2,n · · · sm,1 sm,2 1 · · · sm,n 1 C C C C C C A Call Me Maybe Esh Esh listened to call me maybe once… ⇡ 0 B B B B B B B B B @ u1 u2 ... ... ... um 1 C C C C C C C C C A t1 t2 · · · · · · · · · tn⇡ 0 B B B B B B B B B @ u1 u2 ... ... ... um 1 C C C C C C C C C A t1 t2 · · · · · · · · · tn
- 19. Matrix Completion iswell studied … Start with random vectors around the origin. Run alternating least squares or gradient descent or stochastic gradient descent… All this is Hadoopable™. Users 8 >>>>>>< >>>>>>: 0 B B B B B B @ Song Plays z }| { s1,1 s1,2 14 · · · s1,n s2,1 s2,2 2 · · · s2,n · · · sm,1 sm,2 1 · · · sm,n 1 C C C C C C A Users 8 >>>>>>< >>>>>>: 0 B B B B B B @ Song Plays z }| { s1,1 s1,2 14 · · · s1,n s2,1 s2,2 2 · · · s2,n · · · sm,1 sm,2 1 · · · sm,n 1 C C C C C C A Call Me Maybe Esh Esh listened to call me maybe once… ⇡ 0 B B B B B B B B B @ u1 u2 ... ... ... um 1 C C C C C C C C C A t1 t2 · · · · · · · · · tn⇡ 0 B B B B B B B B B @ u1 u2 ... ... ... um 1 C C C C C C C C C A t1 t2 · · · · · · · · · tn
- 20. 30 Million Songs… WhatToPlay? 75 Million People … 1 Person Every 3 Secs…
- 21.
- 22. Language Models • Languagemodels work well too. For example, a playlist could be considered as a document and you could learn the latent vectors for tracks (words). • Then represent a User as a linear combination of their Tracks.
- 23. word2vec Words with similarcontexts have similar meaning
- 24.
- 25.
- 26. word2vec Target Words andCorresponding Contexts shining bright trees dark green stars 61 50 10 30 1 sun 71 60 5 2 0 cucumber 2 1 15 3 40
- 27. word2vec Playlists CPU Vectors ReadGetVectors & Update
- 28. Vectors are awesome! •Uniquefingerprint for every users, tracks, albums, artists & even playlists in the same space. •Similarity is easily computable. Euclidean Distance or Cosine Similarity.
- 29. Approximate Nearest Neighbors •Fastapproximate nearest neighbor search. • Locality Sensitive Hashing • https://github.com/spotify/annoy
- 30. Vectors are greatfor Infrastructure too… •Machine Learning can be decomposed & abstracted away. •A Lambda Architecture involving Machine Learning becomes eas(ier). •Platforms for Personalization become possible….
- 31. The Record Store… TheList Maker … How do you scale this?
- 32. Tools of thetrade • Build models in Python. (NumPy, SciPy ) • Jobs in Scalding + Luigi ( https://github.com/spotify/luigi ) • Storm for real time. • In house RPC for serving requests.
- 33. Storm 101 • RealtimeStream Processing. • Like Hadoop but easier. • Fault tolerant. • Java, Clojure (yay!) and more!
- 34. Storm @ Spotify •Major users are Ads & Personalization! • Everyteam manages its own cluster. For personalization, we have a 12 node cluster. • Relatively a new tech, compared to Hadoop™.
- 35. So why Storm? •Hadoop is slowwww. Daily UserVector jobs takes ~ 16 hours to run. Small Data FTW! • New Users are important; they need a friend! • What moment are you in? Gym, Running etc?.
- 36. Getting Data AcrossThe Globe
- 37.
- 38. HDFS Kafka Pipeline … User Listens Playlists Realtime Listens Spout UserVector Generation Job Latent Vector Models Track, Artist, Album Vectors
- 39. HDFS Kafka Pipeline … User Listens Playlists Realtime Listens Spout UserVector Generation Job Latent Vector Models Track, Artist, Album Vectors Compressed Listening History Bolts Cassandra Cassandra
- 40. HDFS Kafka Pipeline + Platform User Listens Playlists RealtimeListens Spout User Vector Generation Job Latent Vector Models Track, Artist, Album Vectors Compressed Listening History Bolts Cassandra Cassandra Backend Systems •Top Albums •Top Tracks •Top Playlists
- 41. Discover New User •Goingfrom two weeks of no recommendations to recommendations as soon as a user plays a track. •Successful A/B test •First team to build a production ready personalization feature using Storm.
- 42. Lessons Learnt … •Boring technology works well. Complicated Storm Topology = Bad. (Dan Mckinley) • Storm is nice. Would have preferred reusing batch Scalding Code. Maybe Spark Streaming? • Grow your API from one use case to another. Don’t solve for everything at one time.
- 43. Join the band! •Machine Learning, Data & Backend Gigs. • Now touring in New York, Boston & Stockholm! • https://www.spotify.com/jobs/
- 44.