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.

Music Recommendations at Scale with Spark

37,019 views

Published on

Spotify uses a range of Machine Learning models to power its music recommendation features including the Discover page, Radio, and Related Artists. Due to the iterative nature of these models they are a natural fit to the Spark computation paradigm and suffer from the IO overhead incurred by Hadoop. In this talk, I review the ALS algorithm for Matrix Factorization with implicit feedback data and how we’ve scaled it up to handle 100s of Billions of data points using Scala, Breeze, and Spark.

Published in: Software, Technology, Education
  • Hello! Get Your Professional Job-Winning Resume Here - Check our website! https://vk.cc/818RFv
       Reply 
    Are you sure you want to  Yes  No
    Your message goes here

Music Recommendations at Scale with Spark

  1. June 27, 2014 Music Recommendations at Scale with Spark Chris Johnson @MrChrisJohnson
  2. Who am I?? •Chris Johnson – Machine Learning guy from NYC – Focused on music recommendations – Formerly a PhD student at UT Austin
  3. 3 Recommendations at Spotify ! • Discover (personalized recommendations) • Radio • Related Artists • Now Playing
  4. How can we find good recommendations? ! • Manual Curation ! ! ! • Manually Tag Attributes ! ! • Audio Content, Metadata, Text Analysis ! ! • Collaborative Filtering 4
  5. How can we find good recommendations? ! • Manual Curation ! ! ! • Manually Tag Attributes ! ! • Audio Content, Metadata, Text Analysis ! ! • Collaborative Filtering 5
  6. Collaborative Filtering - “The Netflix Prize” 6
  7. Collaborative Filtering 7 Hey, I like tracks P, Q, R, S! Well, I like tracks Q, R, S, T! Then you should check out track P! Nice! Btw try track T! Image via Erik Bernhardsson
  8. Section name 8
  9. Explicit Matrix Factorization 9 Movies Users Chris Inception •Users explicitly rate a subset of the movie catalog •Goal: predict how users will rate new movies
  10. • = bias for user • = bias for item • = regularization parameter Explicit Matrix Factorization 10 Chris Inception ? 3 5 ? 1 ? ? 1 2 ? 3 2 ? ? ? 5 5 2 ? 4 •Approximate ratings matrix by the product of low- dimensional user and movie matrices •Minimize RMSE (root mean squared error) • = user rating for movie • = user latent factor vector • = item latent factor vector X YUsers Movies
  11. Implicit Matrix Factorization 11 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 1 1 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 1 •Instead of explicit ratings use binary labels – 1 = streamed, 0 = never streamed •Minimize weighted RMSE (root mean squared error) using a function of total streams as weights • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • =i tem latent factor vector X YUsers Songs
  12. Alternating Least Squares (ALS) 12 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 1 1 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 1 •Instead of explicit ratings use binary labels – 1 = streamed, 0 = never streamed •Minimize weighted RMSE (root mean squared error) using a function of total streams as weights • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • =i tem latent factor vector X YUsers Songs Fix songs
  13. Alternating Least Squares (ALS) 13 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 1 1 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 1 •Instead of explicit ratings use binary labels – 1 = streamed, 0 = never streamed •Minimize weighted RMSE (root mean squared error) using a function of total streams as weights • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • =i tem latent factor vector X YUsers Songs Fix songs Solve for users
  14. Alternating Least Squares (ALS) 14 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 1 1 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 1 •Instead of explicit ratings use binary labels – 1 = streamed, 0 = never streamed •Minimize weighted RMSE (root mean squared error) using a function of total streams as weights • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • =i tem latent factor vector X YUsers Songs Fix users
  15. Alternating Least Squares (ALS) 15 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 1 1 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 1 •Instead of explicit ratings use binary labels – 1 = streamed, 0 = never streamed •Minimize weighted RMSE (root mean squared error) using a function of total streams as weights • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • =i tem latent factor vector X YUsers Songs Solve for songs Fix users
  16. Alternating Least Squares (ALS) 16 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 1 1 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 1 •Instead of explicit ratings use binary labels – 1 = streamed, 0 = never streamed •Minimize weighted RMSE (root mean squared error) using a function of total streams as weights • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • =i tem latent factor vector X YUsers Songs Solve for songs Fix users Repeat until convergence…
  17. Alternating Least Squares (ALS) 17 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 1 0 0 0 1 1 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 0 1 0 0 1 •Instead of explicit ratings use binary labels – 1 = streamed, 0 = never streamed •Minimize weighted RMSE (root mean squared error) using a function of total streams as weights • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • =i tem latent factor vector X YUsers Songs Solve for songs Fix users Repeat until convergence…
  18. 18 Alternating Least Squares code: https://github.com/MrChrisJohnson/implicitMF
  19. Section name 19
  20. Scaling up Implicit Matrix Factorization with Hadoop 20
  21. Hadoop at Spotify 2009 21
  22. Hadoop at Spotify 2014 22 700 Nodes in our London data center
  23. Implicit Matrix Factorization with Hadoop 23 Reduce stepMap step u % K = 0 i % L = 0 u % K = 0 i % L = 1 ... u % K = 0 i % L = L-1 u % K = 1 i % L = 0 u % K = 1 i % L = 1 ... ... ... ... ... ... u % K = K-1 i % L = 0 ... ... u % K = K-1 i % L = L-1 item vectors item%L=0 item vectors item%L=1 item vectors i % L = L-1 user vectors u % K = 0 user vectors u % K = 1 user vectors u % K = K-1 all log entries u % K = 1 i % L = 1 u % K = 0 u % K = 1 u % K = K-1 Figure via Erik Bernhardsson
  24. Implicit Matrix Factorization with Hadoop 24 One map task Distributed cache: All user vectors where u % K = x Distributed cache: All item vectors where i % L = y Mapper Emit contributions Map input: tuples (u, i, count) where u % K = x and i % L = y Reducer New vector! Figure via Erik Bernhardsson
  25. Hadoop suffers from I/O overhead 25 IO Bottleneck
  26. Spark to the rescue!! 26 Vs http://www.slideshare.net/Hadoop_Summit/spark-and-shark Spark Hadoop
  27. Section name 27
  28. 28 ratings user vectors item vectors First Attempt (broadcast everything) worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 • For each iteration: 1. Compute YtY over item vectors and broadcast 2. Broadcast item vectors 3. Group ratings by user 4. Solve for optimal user vector
  29. 29 ratings user vectors item vectors First Attempt (broadcast everything) worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY • For each iteration: 1. Compute YtY over item vectors and broadcast 2. Broadcast item vectors 3. Group ratings by user 4. Solve for optimal user vector
  30. First Attempt (broadcast everything) 30 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY • For each iteration: 1. Compute YtY over item vectors and broadcast 2. Broadcast item vectors 3. Group ratings by user 4. Solve for optimal user vector
  31. 31 ratings user vectors item vectors First Attempt (broadcast everything) worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY • For each iteration: 1. Compute YtY over item vectors and broadcast 2. Broadcast item vectors 3. Group ratings by user 4. Solve for optimal user vector
  32. First Attempt (broadcast everything) 32 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 • For each iteration: 1. Compute YtY over item vectors and broadcast 2. Broadcast item vectors 3. Group ratings by user 4. Solve for optimal user vector
  33. First Attempt (broadcast everything) 33
  34. First Attempt (broadcast everything) 34 •Cons: – Unnecessarily shuffling all data across wire each iteration. – Not caching ratings data – Unnecessarily sending a full copy of user/item vectors to all workers.
  35. Second Attempt (full gridify) 35 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 •Group ratings matrix into K x L, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector send a copy to each rating block in the item % L column 3. Compute intermediate terms for each block (partition) 4. Group by user, aggregate intermediate terms, and solve for optimal user vector
  36. Second Attempt (full gridify) 36 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 •Group ratings matrix into K x L, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector send a copy to each rating block in the item % L column 3. Compute intermediate terms for each block (partition) 4. Group by user, aggregate intermediate terms, and solve for optimal user vector
  37. Second Attempt (full gridify) 37 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 •Group ratings matrix into K x L, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector send a copy to each rating block in the item % L column 3. Compute intermediate terms for each block (partition) 4. Group by user, aggregate intermediate terms, and solve for optimal user vector
  38. Second Attempt (full gridify) 38 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY •Group ratings matrix into K x L, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector send a copy to each rating block in the item % L column 3. Compute intermediate terms for each block (partition) 4. Group by user, aggregate intermediate terms, and solve for optimal user vector
  39. Second Attempt (full gridify) 39 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY •Group ratings matrix into K x L, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector send a copy to each rating block in the item % L column 3. Compute intermediate terms for each block (partition) 4. Group by user, aggregate intermediate terms, and solve for optimal user vector
  40. Second Attempt (full gridify) 40 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY •Group ratings matrix into K x L, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector send a copy to each rating block in the item % L column 3. Compute intermediate terms for each block (partition) 4. Group by user, aggregate intermediate terms, and solve for optimal user vector
  41. Second Attempt (full gridify) 41 ratings user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 •Group ratings matrix into K x L, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector send a copy to each rating block in the item % L column 3. Compute intermediate terms for each block (partition) 4. Group by user, aggregate intermediate terms, and solve for optimal user vector
  42. Second Attempt 42
  43. Second Attempt 43 •Pros – Ratings get cached and never shuffled – Each partition only requires a subset of item (or user) vectors in memory each iteration – Potentially requires less local memory than a “half gridify” scheme •Cons - Sending lots of intermediate data over wire each iteration in order to aggregate and solve for optimal vectors - More IO overhead than a “half gridify” scheme
  44. Third Attempt (half gridify) 44 ratings user vectors item vectors •Partition ratings matrix into K user (row) and item (column) blocks, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector, send a copy to each user rating partition that requires it (potentially all partitions) 3. Each partition aggregates intermediate terms and solves for optimal user vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6
  45. Third Attempt (half gridify) 45 ratings user vectors item vectors •Partition ratings matrix into K user (row) and item (column) blocks, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector, send a copy to each user rating partition that requires it (potentially all partitions) 3. Each partition aggregates intermediate terms and solves for optimal user vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6
  46. Third Attempt (half gridify) 46 ratings user vectors item vectors •Partition ratings matrix into K user (row) and item (column) blocks, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector, send a copy to each user rating partition that requires it (potentially all partitions) 3. Each partition aggregates intermediate terms and solves for optimal user vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6
  47. Third Attempt (half gridify) 47 ratings user vectors item vectors •Partition ratings matrix into K user (row) and item (column) blocks, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector, send a copy to each user rating partition that requires it (potentially all partitions) 3. Each partition aggregates intermediate terms and solves for optimal user vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY
  48. Third Attempt (half gridify) 48 ratings user vectors item vectors •Partition ratings matrix into K user (row) and item (column) blocks, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector, send a copy to each user rating partition that requires it (potentially all partitions) 3. Each partition aggregates intermediate terms and solves for optimal user vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY
  49. Third Attempt (half gridify) 49 ratings user vectors item vectors •Partition ratings matrix into K user (row) and item (column) blocks, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector, send a copy to each user rating partition that requires it (potentially all partitions) 3. Each partition aggregates intermediate terms and solves for optimal user vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY
  50. Third Attempt (half gridify) 50 ratings user vectors item vectors •Partition ratings matrix into K user (row) and item (column) blocks, partition, and cache •For each iteration: 1. Compute YtY over item vectors and broadcast 2. For each item vector, send a copy to each user rating partition that requires it (potentially all partitions) 3. Each partition aggregates intermediate terms and solves for optimal user vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY Note that we removed the extra shuffle from the full gridify approach.
  51. 51 Third Attempt (half gridify) •Pros – Ratings get cached and never shuffled – Once item vectors are joined with ratings partitions each partition has enough information to solve optimal user vectors without any additional shuffling/aggregation (which occurs with the “full gridify” scheme) •Cons - Each partition could potentially require a copy of each item vector (which may not all fit in memory) - Potentially requires more local memory than “full gridify” scheme Actual MLlib code!
  52. ALS Running Times 52 Hadoop Spark (full gridify) Spark (half gridify) 10 hours 3.5 hours 1.5 hours •Dataset consisting of Spotify streaming data for 4 Million users and 500k artists -Note: full dataset consists of 40M users and 20M songs but we haven’t yet successfully run with Spark •All jobs run using 40 latent factors •Spark jobs used 200 executors with 20G containers •Hadoop job used 1k mappers, 300 reducers
  53. ALS Running Times 53 ALS runtime numbers via @evansparks using Spark version 0.8.0
  54. Section name 54
  55. Random Learnings 55 •PairRDDFunctions are your friend!
  56. Random Learnings 56 •Kryo serialization faster than java serialization but may require you to write and/or register your own serializers
  57. Random Learnings 57 •Kryo serialization faster than java serialization but may require you to write and/or register your own serializers
  58. Random Learnings 58 •Running with larger datasets often results in failed executors and job never fully recovers
  59. Section name 59 Fin
  60. Section name 60
  61. Section name 61
  62. Section name 62
  63. Section name 63
  64. Section name 64
  65. Section name 65

×