Upload
chris-johnson
View
5.384
Download
1
Tags:
Embed Size (px)
DESCRIPTION
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.
Citation preview
June 27, 2014
Music Recommendations at Scale with Spark
Chris Johnson @MrChrisJohnson
Who am I??•Chris Johnson
– Machine Learning guy from NYC – Focused on music recommendations – Formerly a PhD student at UT Austin
3Recommendations at Spotify !
• Discover (personalized recommendations) • Radio • Related Artists • Now Playing
How can we find good recommendations?
!•Manual Curation !!!
•Manually Tag Attributes !!• Audio Content,
Metadata, Text Analysis !!• Collaborative Filtering
4
How can we find good recommendations?
!•Manual Curation !!!
•Manually Tag Attributes !!• Audio Content,
Metadata, Text Analysis !!• Collaborative Filtering
5
Collaborative Filtering - “The Netflix Prize” 6
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
Section name 8
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
• = bias for user • = bias for item • = regularization parameter
Explicit Matrix Factorization 10
ChrisInception
? 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
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
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
SongsFix songs
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
SongsFix songs
Solve for users
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
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
SongsSolve for songs
Fix users
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
SongsSolve for songs
Fix users
Repeat until convergence…
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
SongsSolve for songs
Fix users
Repeat until convergence…
18Alternating Least Squares
code: https://github.com/MrChrisJohnson/implicitMF
Section name 19
Scaling up Implicit Matrix Factorization with Hadoop
20
Hadoop at Spotify 2009 21
Hadoop at Spotify 2014 22
700 Nodes in our London data center
Implicit Matrix Factorization with Hadoop 23
Reduce stepMap step
u % K = 0i % L = 0
u % K = 0i % L = 1 ... u % K = 0
i % L = L-1
u % K = 1i % L = 0
u % K = 1i % L = 1 ... ...
... ... ... ...
u % K = K-1i % L = 0 ... ... u % K = K-1
i % L = L-1
item vectorsitem%L=0
item vectorsitem%L=1
item vectorsi % L = L-1
user vectorsu % K = 0
user vectorsu % K = 1
user vectorsu % K = K-1
all log entriesu % K = 1i % L = 1
u % K = 0
u % K = 1
u % K = K-1
Figure via Erik Bernhardsson
Implicit Matrix Factorization with Hadoop 24
One map taskDistributed
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
andi % L = y
Reducer New vector!
Figure via Erik Bernhardsson
Hadoop suffers from I/O overhead 25
IO Bottleneck
Spark to the rescue!! 26
Vs
http://www.slideshare.net/Hadoop_Summit/spark-and-shark
Spark
Hadoop
Section name 27
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
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
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
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
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
First Attempt (broadcast everything) 33
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.
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
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
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
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
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
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
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
Second Attempt 42
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
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
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
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
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
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
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
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.
51Third 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!
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
ALS Running Times 53
ALS runtime numbers via @evansparks using Spark version 0.8.0
Section name 54
Random Learnings 55
•PairRDDFunctions are your friend!
Random Learnings 56
•Kryo serialization faster than java serialization but may require you to write and/or register your own serializers
Random Learnings 57
•Kryo serialization faster than java serialization but may require you to write and/or register your own serializers
Random Learnings 58
•Running with larger datasets often results in failed executors and job never fully recovers
Section name 59
Fin
Section name 60
Section name 61
Section name 62
Section name 63
Section name 64
Section name 65