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.

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