SlideShare utilise les cookies pour améliorer les fonctionnalités et les performances, et également pour vous montrer des publicités pertinentes. Si vous continuez à naviguer sur ce site, vous acceptez l’utilisation de cookies. Consultez nos Conditions d’utilisation et notre Politique de confidentialité.

SlideShare utilise les cookies pour améliorer les fonctionnalités et les performances, et également pour vous montrer des publicités pertinentes. Si vous continuez à naviguer sur ce site, vous acceptez l’utilisation de cookies. Consultez notre Politique de confidentialité et nos Conditions d’utilisation pour en savoir plus.

Ce diaporama a bien été signalé.

Vous avez aimé cette présentation ? Partagez !

- Recommender system introduction by Liang Xiang 22820 views
- Intro to Factorization Machines by Pavel Kalaidin 4631 views
- Neighbor methods vs matrix factoriz... by Domonkos Tikk 5364 views
- Matrix factorization by rubyyc 1072 views
- Factorization Machines with libFM by Liangjie Hong 14827 views
- Nonnegative Matrix Factorization by Tatsuya Yokota 4750 views

Spotify uses a range of Machine Learning models to power its music recommendation features including the Discover page and Radio. Due to the iterative nature of training these models they suffer from IO overhead of Hadoop and are a natural fit to the Spark programming paradigm. In this talk I will present both the right way as well as the wrong way to implement collaborative filtering models with Spark. Additionally, I will deep dive into how Matrix Factorization is implemented in the MLlib library.

Aucun téléchargement

Nombre de vues

45 320

Sur SlideShare

0

Issues des intégrations

0

Intégrations

6 113

Partages

0

Téléchargements

1 660

Commentaires

60

J’aime

216

Aucune remarque pour cette diapositive

- 1. May 9, 2014 Collaborative Filtering with Spark Chris Johnson @MrChrisJohnson Friday, May 9, 14
- 2. Who am I?? •Chris Johnson – Machine Learning guy from NYC – Focused on music recommendations – Formerly a graduate student at UT Austin Friday, May 9, 14
- 3. 3 What is MLlib? Algorithms: • classification: logistic regression, linear support vector machine (SVM), naive bayes • regression: generalized linear regression • clustering: k-means • decomposition: singular value decomposition (SVD), principle component analysis (PCA • collaborative filtering: alternating least squares (ALS) http://spark.apache.org/docs/0.9.0/mllib-guide.html Friday, May 9, 14
- 4. 4 What is MLlib? Algorithms: • classification: logistic regression, linear support vector machine (SVM), naive bayes • regression: generalized linear regression • clustering: k-means • decomposition: singular value decomposition (SVD), principle component analysis (PCA • collaborative filtering: alternating least squares (ALS) http://spark.apache.org/docs/0.9.0/mllib-guide.html Friday, May 9, 14
- 5. Collaborative Filtering - “The Netflix Prize” 5 Friday, May 9, 14
- 6. Collaborative Filtering 6 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 Friday, May 9, 14
- 7. 7 Collaborative Filtering at Spotify • Discover (personalized recommendations) • Radio • Related Artists • Now Playing Friday, May 9, 14
- 8. Section name 8 Friday, May 9, 14
- 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 Friday, May 9, 14
- 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 Y Friday, May 9, 14
- 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 •Replace Stream counts with binary labels – 1 = streamed, 0 = never streamed •Minimize weighted RMSE (root mean squared error) using a function of stream counts 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 Y Friday, May 9, 14
- 12. Alternating Least Squares 12 • Initialize user and item vectors to random noise • Fix item vectors and solve for optimal user vectors – Take the derivative of loss function with respect to user’s vector, set equal to 0, and solve – Results in a system of linear equations with closed form solution! • Fix user vectors and solve for optimal item vectors • Repeat until convergence code: https://github.com/MrChrisJohnson/implicitMF Friday, May 9, 14
- 13. Alternating Least Squares 13 • Note that: • Then, we can pre-compute once per iteration – and only contain non-zero elements for tracks that the user streamed – Using sparse matrix operations we can then compute each user’s vector efficiently in time where is the number of tracks the user streamed code: https://github.com/MrChrisJohnson/implicitMF Friday, May 9, 14
- 14. 14 Alternating Least Squares code: https://github.com/MrChrisJohnson/implicitMF Friday, May 9, 14
- 15. Section name 15 Friday, May 9, 14
- 16. Scaling up Implicit Matrix Factorization with Hadoop 16 Friday, May 9, 14
- 17. Hadoop at Spotify 2009 17 Friday, May 9, 14
- 18. Hadoop at Spotify 2014 18 700 Nodes in our London data center Friday, May 9, 14
- 19. Implicit Matrix Factorization with Hadoop 19 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 Friday, May 9, 14
- 20. Implicit Matrix Factorization with Hadoop 20 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 Friday, May 9, 14
- 21. Hadoop suffers from I/O overhead 21 IO Bottleneck Friday, May 9, 14
- 22. Spark to the rescue!! 22 Vs http://www.slideshare.net/Hadoop_Summit/spark-and-shark Spark Hadoop Friday, May 9, 14
- 23. Section name 23 Friday, May 9, 14
- 24. First Attempt 24 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • For each iteration: – Compute YtY over item vectors and broadcast – Join user vectors along with all ratings for that user and all item vectors for which the user rated the item – Sum up YtCuIY and YtCuPu and solve for optimal user vectors Friday, May 9, 14
- 25. First Attempt 25 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • For each iteration: – Compute YtY over item vectors and broadcast – Join user vectors along with all ratings for that user and all item vectors for which the user rated the item – Sum up YtCuIY and YtCuPu and solve for optimal user vectors node 2 node 3 node 4 node 5 Friday, May 9, 14
- 26. First Attempt 26 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • For each iteration: – Compute YtY over item vectors and broadcast – Join user vectors along with all ratings for that user and all item vectors for which the user rated the item – Sum up YtCuIY and YtCuPu and solve for optimal user vectors node 2 node 3 node 4 node 5 Friday, May 9, 14
- 27. First Attempt 27 Friday, May 9, 14
- 28. First Attempt 28 •Issues: –Unnecessarily sending multiple copies of item vector to each node –Unnecessarily shuffling data across cluster at each iteration –Not taking advantage of Spark’s in memory capabilities! Friday, May 9, 14
- 29. Second Attempt 29 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • For each iteration: – Compute YtY over item vectors and broadcast – Group ratings matrix into blocks, and join blocks with necessary user and item vectors (to avoid multiple item vector copies at each node) – Sum up YtCuIY and YtCuPu and solve for optimal user vectors Friday, May 9, 14
- 30. Second Attempt 30 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • For each iteration: – Compute YtY over item vectors and broadcast – Group ratings matrix into blocks, and join blocks with necessary user and item vectors (to avoid multiple item vector copies at each node) – Sum up YtCuIY and YtCuPu and solve for optimal user vectors Friday, May 9, 14
- 31. Second Attempt 31 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • For each iteration: – Compute YtY over item vectors and broadcast – Group ratings matrix into blocks, and join blocks with necessary user and item vectors (to avoid multiple item vector copies at each node) – Sum up YtCuIY and YtCuPu and solve for optimal user vectors Friday, May 9, 14
- 32. Second Attempt 32 Friday, May 9, 14
- 33. Second Attempt 33 Friday, May 9, 14
- 34. Second Attempt 34 •Issues: –Still Unnecessarily shuffling data across cluster at each iteration –Still not taking advantage of Spark’s in memory capabilities! Friday, May 9, 14
- 35. So, what are we missing?... 35 •Partitioner: Defines how the elements in a key-value pair RDD are partitioned across the cluster. node 1 node 2 node 3 node 4 node 5 node 6 user vectors partition 1 partition 2 partition 3 Friday, May 9, 14
- 36. So, what are we missing?... 36 •partitionBy(partitioner): Partitions all elements of the same key to the same node in the cluster, as defined by the partitioner. node 1 node 2 node 3 node 4 node 5 node 6 user vectors partition 1 partition 2 partition 3 Friday, May 9, 14
- 37. So, what are we missing?... 37 •mapPartitions(func): Similar to map, but runs separately on each partition (block) of the RDD, so func must be of type Iterator[T] => Iterator[U] when running on an RDD of type T. node 1 node 2 node 3 node 4 node 5 node 6 user vectors partition 1 partition 2 partition 3 function() function() function() Friday, May 9, 14
- 38. So, what are we missing?... 38 • persist(storageLevel): Set this RDD's storage level to persist (cache) its values across operations after the first time it is computed. node 1 node 2 node 3 node 4 node 5 node 6 user vectors partition 1 partition 2 partition 3 Friday, May 9, 14
- 39. Third Attempt 39 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • Partition ratings matrix, user vectors, and item vectors by user and item blocks and cache partitions in memory • Build InLink and OutLink mappings for users and items – InLink Mapping: Includes the user IDs and vectors for a given block along with the ratings for each user in this block – OutLink Mapping: Includes the item IDs and vectors for a given block along with a list of destination blocks for which to send these vectors • For each iteration: – Compute YtY over item vectors and broadcast – On each item block, use the OutLink mapping to send item vectors to the necessary user blocks – On each user block, use the InLink mapping along with the joined item vectors to update vectors Friday, May 9, 14
- 40. Third Attempt 40 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • Partition ratings matrix, user vectors, and item vectors by user and item blocks and cache partitions in memory • Build InLink and OutLink mappings for users and items – InLink Mapping: Includes the user IDs and vectors for a given block along with the ratings for each user in this block – OutLink Mapping: Includes the item IDs and vectors for a given block along with a list of destination blocks for which to send these vectors • For each iteration: – Compute YtY over item vectors and broadcast – On each item block, use the OutLink mapping to send item vectors to the necessary user blocks – On each user block, use the InLink mapping along with the joined item vectors to update vectors Friday, May 9, 14
- 41. Third Attempt 41 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • Partition ratings matrix, user vectors, and item vectors by user and item blocks and cache partitions in memory • Build InLink and OutLink mappings for users and items – InLink Mapping: Includes the user IDs and vectors for a given block along with the ratings for each user in this block – OutLink Mapping: Includes the item IDs and vectors for a given block along with a list of destination blocks for which to send these vectors • For each iteration: – Compute YtY over item vectors and broadcast – On each item block, use the OutLink mapping to send item vectors to the necessary user blocks – On each user block, use the InLink mapping along with the joined item vectors to update vectors Friday, May 9, 14
- 42. Third Attempt 42 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • Partition ratings matrix, user vectors, and item vectors by user and item blocks and cache partitions in memory • Build InLink and OutLink mappings for users and items – InLink Mapping: Includes the user IDs and vectors for a given block along with the ratings for each user in this block – OutLink Mapping: Includes the item IDs and vectors for a given block along with a list of destination blocks for which to send these vectors • For each iteration: – Compute YtY over item vectors and broadcast – On each item block, use the OutLink mapping to send item vectors to the necessary user blocks – On each user block, use the InLink mapping along with the joined item vectors to update vectors Friday, May 9, 14
- 43. Third Attempt 43 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • Partition ratings matrix, user vectors, and item vectors by user and item blocks and cache partitions in memory • Build InLink and OutLink mappings for users and items – InLink Mapping: Includes the user IDs and vectors for a given block along with the ratings for each user in this block – OutLink Mapping: Includes the item IDs and vectors for a given block along with a list of destination blocks for which to send these vectors • For each iteration: – Compute YtY over item vectors and broadcast – On each item block, use the OutLink mapping to send item vectors to the necessary user blocks – On each user block, use the InLink mapping along with the joined item vectors to update vectors Friday, May 9, 14
- 44. Third Attempt 44 ratings user vectors item vectors node 1 node 2 node 3 node 4 node 5 node 6 • Partition ratings matrix, user vectors, and item vectors by user and item blocks and cache partitions in memory • Build InLink and OutLink mappings for users and items – InLink Mapping: Includes the user IDs and vectors for a given block along with the ratings for each user in this block – OutLink Mapping: Includes the item IDs and vectors for a given block along with a list of destination blocks for which to send these vectors • For each iteration: – Compute YtY over item vectors and broadcast – On each item block, use the OutLink mapping to send item vectors to the necessary user blocks – On each user block, use the InLink mapping along with the joined item vectors to update vectors Friday, May 9, 14
- 45. Third attempt 45 Friday, May 9, 14
- 46. Third attempt 46 Friday, May 9, 14
- 47. Third attempt 47 Friday, May 9, 14
- 48. ALS Running Times 48 Via Xiangrui Meng (Databricks) http://stanford.edu/~rezab/sparkworkshop/slides/xiangrui.pdf Friday, May 9, 14
- 49. Section name 49 Fin Friday, May 9, 14
- 50. Section name 50 Friday, May 9, 14
- 51. Section name 51 Friday, May 9, 14
- 52. Section name 52 Friday, May 9, 14
- 53. Section name 53 Friday, May 9, 14
- 54. Section name 54 Friday, May 9, 14
- 55. Section name 55 Friday, May 9, 14
- 56. Section name 56 Friday, May 9, 14

Aucun clipboard public n’a été trouvé avec cette diapositive

Il semblerait que vous ayez déjà ajouté cette diapositive à .

Créer un clipboard

Identifiez-vous pour voir les commentaires