SlideShare une entreprise Scribd logo
1  sur  50
Télécharger pour lire hors ligne
January 6, 2015
Scala Data Pipelines for
Music Recommendations
Chris Johnson
@MrChrisJohnson
Who am I??
•Chris Johnson
– Machine Learning guy from NYC
– Focused on music recommendations
– Formerly a PhD student at UTAustin
Spotify in Numbers 3
•Started in 2006, now available in 58 markets
•50+ million active users, 15 million paying subscribers
•30+ million songs, 20,000 new songs added per day
•1.5 billion playlists
•1 TB user data logged per day
•900 node Hadoop cluster
•10,000+ Hadoop jobs run every day
4
Music Recommendations at Spotify
•Discover
•Radio
•Related Artists
How can we find good recommendations? 5
•Manual Curation
•Manually Tag Attributes
•Audio Content
•News, Blogs, Text analysis
•Collaborative Filtering
Music Recommendations Data Flow 6
Why ? 7
Why ? 8
Interview Question
The Genre Toplist Problem 9
•Assume we have access to daily log data for all plays on Spotify.
•Goal: Calculate the top 1k artists on for each genre based on total daily plays
{"User": “userA”, "Date": “2015-01-10", "Artist": “Beyonce", "Track": "Halo", "Genres": ["Pop", "R&B", "Soul"]}
{"User": “userB”, "Date": “2015-01-10”, "Artist": "Led Zeppelin”, "Track": "Achilles Last Stand", "Genres": ["Rock",
"Blues Rock", "Hard Rock"]}
……….
Genre Toplists with Python MapReduce 10
11
Scalding is a Scala library that makes it easy to specify Hadoop
MapReduce jobs. Scalding is built on top of Cascading, a Java
library that abstracts away low-level Hadoop details. Scalding is
comparable to Pig, but offers tight integration with Scala, bringing
advantages of Scala to your MapReduce jobs.
-Twitter
Genre Toplists with Scalding 12
Why ? 13
•Data pipeline flows naturally follow the functional paradigm
•Productivity without sacrificing performance
•Active community and ecosystem
-Scalding
-Summingbird
-Algebird
-Spark
-Breeze
•Many data storage solutions integrate well with JVM
-Cassandra
-HBase
-Voldemort
-Datomic
Spotify’s Scalding repository over time 14
Genre Toplists with Scalding 15
sortWithTake doesn’t fully sort 16
•Uses PriorityQueueMonoid from Algebird library
•What is a Monoid??
-Definition: A Set S and a binary operation • : S x S —> S such that
1. Associativity: For all a, b, and c in S the equation

(a • b) • c = a • (b • c) holds
2. Identity Element: There exists an element e in S such that for every
element a in S, the equations e • a = a • e = a hold
•Example: The natural numbers N under the addition operation.

(1 + 2) + 3 = 1 + (2 + 3)

0 + 1 = 1 + 0 = 1
class PriorityQueueMonoid[K](max : Int)(implicit ord :
Ordering[K]) extends Monoid[PriorityQueue[K]]
sortWithTake 17
•Uses PriorityQueueMonoid from Algebird
•PriorityQueue aggregations form a commutative monoid!
1. Associative:

PQ1 = [ (Jay Z, 545), (Miles Davis, 272), …]

PQ2 = [ (Beyonce, 731), (Kurt Vile, 372), …]

PQ3 = [ (Twin Shadow, 87), … ]

PQ1 ++ (PQ2 ++ PQ3) = (PQ1 ++ PQ2) ++ PQ3
2.Commutative: 

PQ1 ++ PQ2 = PQ2 ++ PQ1
3.Identity: 

PQ1 ++ EmptyPQ = PQ1
class PriorityQueueMonoid[K](max : Int)(implicit ord :
Ordering[K]) extends Monoid[PriorityQueue[K]]
sortWithTake 18
•Uses PriorityQueueMonoid from Algebird
•Ok, great observation… but what’s the point of all this!??
-All monoid aggregations and reduces can begin on the Mapper side
and finish on the Reducer side since the order doesn’t matter!
-Scalding implicitly takes care of Mapper side combining and custom
combiner
-Reduces network traffic to reducers
class PriorityQueueMonoid[K](max : Int)(implicit ord :
Ordering[K]) extends Monoid[PriorityQueue[K]]
reduced traffic
Section name 19
How do we store track metadata? 20
•Lots of metadata associated with tracks (100+ columns!)
-artist, album, record label, genres, audio features, …
•Options:
1. Store each track as one long row with many columns
-Sending lots of data over network when you only need 1 or 2 columns
2. Store each column as a separate data source
-Jobs require costly joins, especially when requiring many columns
•Can we do better?..
Apache Parquet to the rescue! 21
•Apache Parquet is a columnar storage format available to any project in the Hadoop
ecosystem, regardless of the choice of data processing framework, data model or
programming language.
•Efficiently read a subset of columns without scanning the entire dataset
•Row group: A logical horizontal partitioning of the data into rows. There is no
physical structure that is guaranteed for a row group. A row group consists of a
column chunk for each column in the dataset.
•Column chunk: A chunk of the data for a particular column. These live in a particular
row group and is guaranteed to be contiguous in the file.
•Predicate push-down: Define predicates (<, >, <=, …) to filter out column chunks or
even full row groups, evaluated at Hadoop InputFormat layer before Avro conversion
Genre Toplists with Scalding + Parquet 22
Driven - job visualization and performance analytics 23
Luigi - data plumbing since 2012 24
•Workflow management framework developed by Spotify
•Python luigi configuration takes care of dependency resolution, job
scheduling, fault tolerance, etc.
•Support for Hive queries, MapReduce jobs, python snippets, Scalding,
Crunch, Spark, and more!
•Like Oozie but without all of the messy XML
https://github.com/spotify/luigi
Luigi 25
Section name 26
So…. back to music recommendations! 27
•Manual Curation
•Manually Tag Attributes
•Audio Content
•News, Blogs, Text analysis
•Collaborative Filtering
Collaborative Filtering
28
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
Implicit Matrix Factorization 29
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
•Aggregate all (user, track) streams into a large matrix
•Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by
minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight
•Why?: Once learned, the top recommendations for a user are the top inner products between
their latent factor vector in X and the track latent factor vectors in Y.
X YUsers
Songs
• = bias for user
• = bias for item
• = regularization parameter
• = 1 if user streamed track else 0
•
• = user latent factor vector
• = item latent factor vector
Alternating Least Squares 30
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
X YUsers
Songs
• = bias for user
• = bias for item
• = regularization parameter
• = 1 if user streamed track else 0
•
• = user latent factor vector
• = item latent factor vector
Fix tracks
•Aggregate all (user, track) streams into a large matrix
•Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by
minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight
•Why?: Once learned, the top recommendations for a user are the top inner products between
their latent factor vector in X and the track latent factor vectors in Y.
31
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
X YUsers
Songs
• = bias for user
• = bias for item
• = regularization parameter
• = 1 if user streamed track else 0
•
• = user latent factor vector
• = item latent factor vector
Fix tracks
Solve for users
•Aggregate all (user, track) streams into a large matrix
•Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by
minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight
•Why?: Once learned, the top recommendations for a user are the top inner products between
their latent factor vector in X and the track latent factor vectors in Y.
Alternating Least Squares
32
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
X YUsers
Songs
• = bias for user
• = bias for item
• = regularization parameter
• = 1 if user streamed track else 0
•
• = user latent factor vector
• = item latent factor vector
Fix users
•Aggregate all (user, track) streams into a large matrix
•Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by
minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight
•Why?: Once learned, the top recommendations for a user are the top inner products between
their latent factor vector in X and the track latent factor vectors in Y.
Alternating Least Squares
33
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
X YUsers
Songs
• = bias for user
• = bias for item
• = regularization parameter
• = 1 if user streamed track else 0
•
• = user latent factor vector
• = item latent factor vector
Fix users
Solve for tracks
•Aggregate all (user, track) streams into a large matrix
•Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by
minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight
•Why?: Once learned, the top recommendations for a user are the top inner products between
their latent factor vector in X and the track latent factor vectors in Y.
Alternating Least Squares
34
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
X YUsers
Songs
• = bias for user
• = bias for item
• = regularization parameter
• = 1 if user streamed track else 0
•
• = user latent factor vector
• = item latent factor vector
Fix users
Solve for tracks
Repeat until convergence…
•Aggregate all (user, track) streams into a large matrix
•Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by
minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight
•Why?: Once learned, the top recommendations for a user are the top inner products between
their latent factor vector in X and the track latent factor vectors in Y.
Alternating Least Squares
35
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
X YUsers
Songs
• = bias for user
• = bias for item
• = regularization parameter
• = 1 if user streamed track else 0
•
• = user latent factor vector
• = item latent factor vector
Fix users
Solve for tracks
Repeat until convergence…
•Aggregate all (user, track) streams into a large matrix
•Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by
minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight
•Why?: Once learned, the top recommendations for a user are the top inner products between
their latent factor vector in X and the track latent factor vectors in Y.
Alternating Least Squares
Matrix Factorization with MapReduce
36
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
Matrix Factorization with MapReduce
37
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
38
•Fast and general purpose cluster computing system
•Provides high-level apis in Java, Scala, and Python
•Takes advantage of in-memory caching to reduce I/O bottleneck of
Hadoop MapReduce
•MLlib: Scalable Machine Learning library packaged with Spark
-Collaborative Filtering and Matrix Factorization
-Classification and Regression
-Clustering
-Optimization Primitives
•Spark Streaming: Real time, scalable, fault-tolerant stream processing
•Spark SQL: allows relational queries expressed in SQL, HiveQL, or
Scala to be executed using Spark
Matrix Factorization with Spark
39
streams user vectors item vectors
worker 1 worker 2 worker 3 worker 4 worker 5 worker 6
•Partition streams matrix into user (row) and item (column) blocks, partition, and cache
-Unlike with the MapReduce implementation, ratings are never shuffled across the network!
•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
Matrix Factorization with Spark
40
user vectors item vectors
worker 1 worker 2 worker 3 worker 4 worker 5 worker 6
•Partition streams matrix into user (row) and item (column) blocks, partition, and cache
-Unlike with the MapReduce implementation, ratings are never shuffled across the network!
•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
streams
Matrix Factorization with Spark
41
user vectors item vectors
•Partition streams matrix into user (row) and item (column) blocks, partition, and cache
-Unlike with the MapReduce implementation, ratings are never shuffled across the network!
•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
streams
Matrix Factorization with Spark
42
user vectors item vectors
worker 1 worker 2 worker 3 worker 4 worker 5 worker 6
YtY YtY YtY YtY YtY YtY
•Partition streams matrix into user (row) and item (column) blocks, partition, and cache
-Unlike with the MapReduce implementation, ratings are never shuffled across the network!
•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
streams
Matrix Factorization with Spark
43
user vectors item vectors
worker 1 worker 2 worker 3 worker 4 worker 5 worker 6
YtY YtY YtY YtY YtY YtY
•Partition streams matrix into user (row) and item (column) blocks, partition, and cache
-Unlike with the MapReduce implementation, ratings are never shuffled across the network!
•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
streams
Matrix Factorization with Spark
44
user vectors item vectors
worker 1 worker 2 worker 3 worker 4 worker 5 worker 6
YtY YtY YtY YtY YtY YtY
•Partition streams matrix into user (row) and item (column) blocks, partition, and cache
-Unlike with the MapReduce implementation, ratings are never shuffled across the network!
•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
streams
45
Vs
http://www.slideshare.net/Hadoop_Summit/spark-and-shark
Matrix Factorization with MapReduce
Matrix Factorization with Spark
Scala Breeze 46
•Native Scala numerical processing library
•Linear Algebra
-Matrix operations
-Operator overloading and syntactic sugar
•Sampling from Probably Distributions
•Numerical Optimization
•Plotting and Visualizations
•Numpy for Scala
Zeppelin + Spark + Parquet for ETL 47
Zeppelin + Spark + Parquet for ETL 48
What should I be worried about? 49
•Multiple “right” ways to do the same thing
•Implicits can make code difficult to navigate
•Learning curve can be tough
•Avoid flattening before a join
•Be aware that Scala default collections are immutable (though mutable
versions are also available)
•Use monoid reduces and aggregations where possible and avoid folds
•Be patient with the compiler
Section name 50
Fin

Contenu connexe

Tendances

Machine learning @ Spotify - Madison Big Data Meetup
Machine learning @ Spotify - Madison Big Data MeetupMachine learning @ Spotify - Madison Big Data Meetup
Machine learning @ Spotify - Madison Big Data MeetupAndy Sloane
 
Scala Data Pipelines @ Spotify
Scala Data Pipelines @ SpotifyScala Data Pipelines @ Spotify
Scala Data Pipelines @ SpotifyNeville Li
 
Personalizing the listening experience
Personalizing the listening experiencePersonalizing the listening experience
Personalizing the listening experienceMounia Lalmas-Roelleke
 
Music recommendations @ MLConf 2014
Music recommendations @ MLConf 2014Music recommendations @ MLConf 2014
Music recommendations @ MLConf 2014Erik Bernhardsson
 
Music Personalization At Spotify
Music Personalization At SpotifyMusic Personalization At Spotify
Music Personalization At SpotifyVidhya Murali
 
Music Personalization : Real time Platforms.
Music Personalization : Real time Platforms.Music Personalization : Real time Platforms.
Music Personalization : Real time Platforms.Esh Vckay
 
Machine Learning and Big Data for Music Discovery at Spotify
Machine Learning and Big Data for Music Discovery at SpotifyMachine Learning and Big Data for Music Discovery at Spotify
Machine Learning and Big Data for Music Discovery at SpotifyChing-Wei Chen
 
Homepage Personalization at Spotify
Homepage Personalization at SpotifyHomepage Personalization at Spotify
Homepage Personalization at SpotifyOguz Semerci
 
Interactive Recommender Systems with Netflix and Spotify
Interactive Recommender Systems with Netflix and SpotifyInteractive Recommender Systems with Netflix and Spotify
Interactive Recommender Systems with Netflix and SpotifyChris Johnson
 
Recommending and Searching (Research @ Spotify)
Recommending and Searching (Research @ Spotify)Recommending and Searching (Research @ Spotify)
Recommending and Searching (Research @ Spotify)Mounia Lalmas-Roelleke
 
ML+Hadoop at NYC Predictive Analytics
ML+Hadoop at NYC Predictive AnalyticsML+Hadoop at NYC Predictive Analytics
ML+Hadoop at NYC Predictive AnalyticsErik Bernhardsson
 
Crafting Recommenders: the Shallow and the Deep of it!
Crafting Recommenders: the Shallow and the Deep of it! Crafting Recommenders: the Shallow and the Deep of it!
Crafting Recommenders: the Shallow and the Deep of it! Sudeep Das, Ph.D.
 
Recommender Systems (Machine Learning Summer School 2014 @ CMU)
Recommender Systems (Machine Learning Summer School 2014 @ CMU)Recommender Systems (Machine Learning Summer School 2014 @ CMU)
Recommender Systems (Machine Learning Summer School 2014 @ CMU)Xavier Amatriain
 
Recommender Systems! @ASAI 2011
Recommender Systems! @ASAI 2011Recommender Systems! @ASAI 2011
Recommender Systems! @ASAI 2011Ernesto Mislej
 
Boston ML - Architecting Recommender Systems
Boston ML - Architecting Recommender SystemsBoston ML - Architecting Recommender Systems
Boston ML - Architecting Recommender SystemsJames Kirk
 
Interactive Recommender Systems
Interactive Recommender SystemsInteractive Recommender Systems
Interactive Recommender SystemsRoelof van Zwol
 
Spotify Discover Weekly: The machine learning behind your music recommendations
Spotify Discover Weekly: The machine learning behind your music recommendationsSpotify Discover Weekly: The machine learning behind your music recommendations
Spotify Discover Weekly: The machine learning behind your music recommendationsSophia Ciocca
 
Deeper Things: How Netflix Leverages Deep Learning in Recommendations and Se...
 Deeper Things: How Netflix Leverages Deep Learning in Recommendations and Se... Deeper Things: How Netflix Leverages Deep Learning in Recommendations and Se...
Deeper Things: How Netflix Leverages Deep Learning in Recommendations and Se...Sudeep Das, Ph.D.
 
[Final]collaborative filtering and recommender systems
[Final]collaborative filtering and recommender systems[Final]collaborative filtering and recommender systems
[Final]collaborative filtering and recommender systemsFalitokiniaina Rabearison
 

Tendances (20)

Machine learning @ Spotify - Madison Big Data Meetup
Machine learning @ Spotify - Madison Big Data MeetupMachine learning @ Spotify - Madison Big Data Meetup
Machine learning @ Spotify - Madison Big Data Meetup
 
Scala Data Pipelines @ Spotify
Scala Data Pipelines @ SpotifyScala Data Pipelines @ Spotify
Scala Data Pipelines @ Spotify
 
Personalizing the listening experience
Personalizing the listening experiencePersonalizing the listening experience
Personalizing the listening experience
 
Music recommendations @ MLConf 2014
Music recommendations @ MLConf 2014Music recommendations @ MLConf 2014
Music recommendations @ MLConf 2014
 
Music Personalization At Spotify
Music Personalization At SpotifyMusic Personalization At Spotify
Music Personalization At Spotify
 
Music Personalization : Real time Platforms.
Music Personalization : Real time Platforms.Music Personalization : Real time Platforms.
Music Personalization : Real time Platforms.
 
Machine Learning and Big Data for Music Discovery at Spotify
Machine Learning and Big Data for Music Discovery at SpotifyMachine Learning and Big Data for Music Discovery at Spotify
Machine Learning and Big Data for Music Discovery at Spotify
 
Homepage Personalization at Spotify
Homepage Personalization at SpotifyHomepage Personalization at Spotify
Homepage Personalization at Spotify
 
Interactive Recommender Systems with Netflix and Spotify
Interactive Recommender Systems with Netflix and SpotifyInteractive Recommender Systems with Netflix and Spotify
Interactive Recommender Systems with Netflix and Spotify
 
Recommending and Searching (Research @ Spotify)
Recommending and Searching (Research @ Spotify)Recommending and Searching (Research @ Spotify)
Recommending and Searching (Research @ Spotify)
 
ML+Hadoop at NYC Predictive Analytics
ML+Hadoop at NYC Predictive AnalyticsML+Hadoop at NYC Predictive Analytics
ML+Hadoop at NYC Predictive Analytics
 
Recommending and searching @ Spotify
Recommending and searching @ SpotifyRecommending and searching @ Spotify
Recommending and searching @ Spotify
 
Crafting Recommenders: the Shallow and the Deep of it!
Crafting Recommenders: the Shallow and the Deep of it! Crafting Recommenders: the Shallow and the Deep of it!
Crafting Recommenders: the Shallow and the Deep of it!
 
Recommender Systems (Machine Learning Summer School 2014 @ CMU)
Recommender Systems (Machine Learning Summer School 2014 @ CMU)Recommender Systems (Machine Learning Summer School 2014 @ CMU)
Recommender Systems (Machine Learning Summer School 2014 @ CMU)
 
Recommender Systems! @ASAI 2011
Recommender Systems! @ASAI 2011Recommender Systems! @ASAI 2011
Recommender Systems! @ASAI 2011
 
Boston ML - Architecting Recommender Systems
Boston ML - Architecting Recommender SystemsBoston ML - Architecting Recommender Systems
Boston ML - Architecting Recommender Systems
 
Interactive Recommender Systems
Interactive Recommender SystemsInteractive Recommender Systems
Interactive Recommender Systems
 
Spotify Discover Weekly: The machine learning behind your music recommendations
Spotify Discover Weekly: The machine learning behind your music recommendationsSpotify Discover Weekly: The machine learning behind your music recommendations
Spotify Discover Weekly: The machine learning behind your music recommendations
 
Deeper Things: How Netflix Leverages Deep Learning in Recommendations and Se...
 Deeper Things: How Netflix Leverages Deep Learning in Recommendations and Se... Deeper Things: How Netflix Leverages Deep Learning in Recommendations and Se...
Deeper Things: How Netflix Leverages Deep Learning in Recommendations and Se...
 
[Final]collaborative filtering and recommender systems
[Final]collaborative filtering and recommender systems[Final]collaborative filtering and recommender systems
[Final]collaborative filtering and recommender systems
 

Similaire à Scala Data Pipelines for Music Recommendations

Collaborative Filtering with Spark
Collaborative Filtering with SparkCollaborative Filtering with Spark
Collaborative Filtering with SparkChris Johnson
 
iliananpappi_mscthesis
iliananpappi_mscthesisiliananpappi_mscthesis
iliananpappi_mscthesisIliana Pappi
 
Random Walk with Restart for Automatic Playlist Continuation and Query-specif...
Random Walk with Restart for Automatic Playlist Continuation and Query-specif...Random Walk with Restart for Automatic Playlist Continuation and Query-specif...
Random Walk with Restart for Automatic Playlist Continuation and Query-specif...Timo van Niedek
 
Approximate nearest neighbor methods and vector models – NYC ML meetup
Approximate nearest neighbor methods and vector models – NYC ML meetupApproximate nearest neighbor methods and vector models – NYC ML meetup
Approximate nearest neighbor methods and vector models – NYC ML meetupErik Bernhardsson
 
Approximate Nearest Neighbors and Vector Models by Erik Bernhardsson
Approximate Nearest Neighbors and Vector Models by Erik BernhardssonApproximate Nearest Neighbors and Vector Models by Erik Bernhardsson
Approximate Nearest Neighbors and Vector Models by Erik BernhardssonHakka Labs
 
Data science-2013-heekim
Data science-2013-heekimData science-2013-heekim
Data science-2013-heekimHaklae Kim
 
A Unified Music Recommender System Using Listening Habits and Semantics of Tags
A Unified Music Recommender System Using Listening Habits and Semantics of TagsA Unified Music Recommender System Using Listening Habits and Semantics of Tags
A Unified Music Recommender System Using Listening Habits and Semantics of Tagsdatasciencekorea
 
Hadoop Tutorial with @techmilind
Hadoop Tutorial with @techmilindHadoop Tutorial with @techmilind
Hadoop Tutorial with @techmilindEMC
 
Modeling and Aggregation of Complex Annotations
Modeling and Aggregation of Complex AnnotationsModeling and Aggregation of Complex Annotations
Modeling and Aggregation of Complex AnnotationsAlexander Braylan
 
A More Scaleable Way of Making Recommendations with MLlib-(Xiangrui Meng, Dat...
A More Scaleable Way of Making Recommendations with MLlib-(Xiangrui Meng, Dat...A More Scaleable Way of Making Recommendations with MLlib-(Xiangrui Meng, Dat...
A More Scaleable Way of Making Recommendations with MLlib-(Xiangrui Meng, Dat...Spark Summit
 
Lecture 7: Data-Intensive Computing for Text Analysis (Fall 2011)
Lecture 7: Data-Intensive Computing for Text Analysis (Fall 2011)Lecture 7: Data-Intensive Computing for Text Analysis (Fall 2011)
Lecture 7: Data-Intensive Computing for Text Analysis (Fall 2011)Matthew Lease
 
streamingalgo88585858585858585pppppp.pptx
streamingalgo88585858585858585pppppp.pptxstreamingalgo88585858585858585pppppp.pptx
streamingalgo88585858585858585pppppp.pptxGopiNathVelivela
 
React Native Performance
React Native Performance React Native Performance
React Native Performance InnerFood
 
Sparkling Random Ferns by P Dendek and M Fedoryszak
Sparkling Random Ferns by  P Dendek and M FedoryszakSparkling Random Ferns by  P Dendek and M Fedoryszak
Sparkling Random Ferns by P Dendek and M FedoryszakSpark Summit
 
A Production Quality Sketching Library for the Analysis of Big Data
A Production Quality Sketching Library for the Analysis of Big DataA Production Quality Sketching Library for the Analysis of Big Data
A Production Quality Sketching Library for the Analysis of Big DataDatabricks
 
Recsys 2018 overview and highlights
Recsys 2018 overview and highlightsRecsys 2018 overview and highlights
Recsys 2018 overview and highlightsSandra Garcia
 

Similaire à Scala Data Pipelines for Music Recommendations (20)

Collaborative Filtering with Spark
Collaborative Filtering with SparkCollaborative Filtering with Spark
Collaborative Filtering with Spark
 
iliananpappi_mscthesis
iliananpappi_mscthesisiliananpappi_mscthesis
iliananpappi_mscthesis
 
Random Walk with Restart for Automatic Playlist Continuation and Query-specif...
Random Walk with Restart for Automatic Playlist Continuation and Query-specif...Random Walk with Restart for Automatic Playlist Continuation and Query-specif...
Random Walk with Restart for Automatic Playlist Continuation and Query-specif...
 
Approximate nearest neighbor methods and vector models – NYC ML meetup
Approximate nearest neighbor methods and vector models – NYC ML meetupApproximate nearest neighbor methods and vector models – NYC ML meetup
Approximate nearest neighbor methods and vector models – NYC ML meetup
 
Approximate Nearest Neighbors and Vector Models by Erik Bernhardsson
Approximate Nearest Neighbors and Vector Models by Erik BernhardssonApproximate Nearest Neighbors and Vector Models by Erik Bernhardsson
Approximate Nearest Neighbors and Vector Models by Erik Bernhardsson
 
Lecture 1 (bce-7)
Lecture   1 (bce-7)Lecture   1 (bce-7)
Lecture 1 (bce-7)
 
Data science-2013-heekim
Data science-2013-heekimData science-2013-heekim
Data science-2013-heekim
 
A Unified Music Recommender System Using Listening Habits and Semantics of Tags
A Unified Music Recommender System Using Listening Habits and Semantics of TagsA Unified Music Recommender System Using Listening Habits and Semantics of Tags
A Unified Music Recommender System Using Listening Habits and Semantics of Tags
 
Hadoop Tutorial with @techmilind
Hadoop Tutorial with @techmilindHadoop Tutorial with @techmilind
Hadoop Tutorial with @techmilind
 
Data Mining Lecture_9.pptx
Data Mining Lecture_9.pptxData Mining Lecture_9.pptx
Data Mining Lecture_9.pptx
 
Modeling and Aggregation of Complex Annotations
Modeling and Aggregation of Complex AnnotationsModeling and Aggregation of Complex Annotations
Modeling and Aggregation of Complex Annotations
 
SVD.ppt
SVD.pptSVD.ppt
SVD.ppt
 
A More Scaleable Way of Making Recommendations with MLlib-(Xiangrui Meng, Dat...
A More Scaleable Way of Making Recommendations with MLlib-(Xiangrui Meng, Dat...A More Scaleable Way of Making Recommendations with MLlib-(Xiangrui Meng, Dat...
A More Scaleable Way of Making Recommendations with MLlib-(Xiangrui Meng, Dat...
 
Lecture 7: Data-Intensive Computing for Text Analysis (Fall 2011)
Lecture 7: Data-Intensive Computing for Text Analysis (Fall 2011)Lecture 7: Data-Intensive Computing for Text Analysis (Fall 2011)
Lecture 7: Data-Intensive Computing for Text Analysis (Fall 2011)
 
streamingalgo88585858585858585pppppp.pptx
streamingalgo88585858585858585pppppp.pptxstreamingalgo88585858585858585pppppp.pptx
streamingalgo88585858585858585pppppp.pptx
 
React Native Performance
React Native Performance React Native Performance
React Native Performance
 
Sparkling Random Ferns by P Dendek and M Fedoryszak
Sparkling Random Ferns by  P Dendek and M FedoryszakSparkling Random Ferns by  P Dendek and M Fedoryszak
Sparkling Random Ferns by P Dendek and M Fedoryszak
 
A Production Quality Sketching Library for the Analysis of Big Data
A Production Quality Sketching Library for the Analysis of Big DataA Production Quality Sketching Library for the Analysis of Big Data
A Production Quality Sketching Library for the Analysis of Big Data
 
Enar short course
Enar short courseEnar short course
Enar short course
 
Recsys 2018 overview and highlights
Recsys 2018 overview and highlightsRecsys 2018 overview and highlights
Recsys 2018 overview and highlights
 

Dernier

SpotFlow: Tracking Method Calls and States at Runtime
SpotFlow: Tracking Method Calls and States at RuntimeSpotFlow: Tracking Method Calls and States at Runtime
SpotFlow: Tracking Method Calls and States at Runtimeandrehoraa
 
Unveiling the Future: Sylius 2.0 New Features
Unveiling the Future: Sylius 2.0 New FeaturesUnveiling the Future: Sylius 2.0 New Features
Unveiling the Future: Sylius 2.0 New FeaturesŁukasz Chruściel
 
SuccessFactors 1H 2024 Release - Sneak-Peek by Deloitte Germany
SuccessFactors 1H 2024 Release - Sneak-Peek by Deloitte GermanySuccessFactors 1H 2024 Release - Sneak-Peek by Deloitte Germany
SuccessFactors 1H 2024 Release - Sneak-Peek by Deloitte GermanyChristoph Pohl
 
Machine Learning Software Engineering Patterns and Their Engineering
Machine Learning Software Engineering Patterns and Their EngineeringMachine Learning Software Engineering Patterns and Their Engineering
Machine Learning Software Engineering Patterns and Their EngineeringHironori Washizaki
 
MYjobs Presentation Django-based project
MYjobs Presentation Django-based projectMYjobs Presentation Django-based project
MYjobs Presentation Django-based projectAnoyGreter
 
Ahmed Motair CV April 2024 (Senior SW Developer)
Ahmed Motair CV April 2024 (Senior SW Developer)Ahmed Motair CV April 2024 (Senior SW Developer)
Ahmed Motair CV April 2024 (Senior SW Developer)Ahmed Mater
 
KnowAPIs-UnknownPerf-jaxMainz-2024 (1).pptx
KnowAPIs-UnknownPerf-jaxMainz-2024 (1).pptxKnowAPIs-UnknownPerf-jaxMainz-2024 (1).pptx
KnowAPIs-UnknownPerf-jaxMainz-2024 (1).pptxTier1 app
 
Intelligent Home Wi-Fi Solutions | ThinkPalm
Intelligent Home Wi-Fi Solutions | ThinkPalmIntelligent Home Wi-Fi Solutions | ThinkPalm
Intelligent Home Wi-Fi Solutions | ThinkPalmSujith Sukumaran
 
SensoDat: Simulation-based Sensor Dataset of Self-driving Cars
SensoDat: Simulation-based Sensor Dataset of Self-driving CarsSensoDat: Simulation-based Sensor Dataset of Self-driving Cars
SensoDat: Simulation-based Sensor Dataset of Self-driving CarsChristian Birchler
 
Taming Distributed Systems: Key Insights from Wix's Large-Scale Experience - ...
Taming Distributed Systems: Key Insights from Wix's Large-Scale Experience - ...Taming Distributed Systems: Key Insights from Wix's Large-Scale Experience - ...
Taming Distributed Systems: Key Insights from Wix's Large-Scale Experience - ...Natan Silnitsky
 
Software Coding for software engineering
Software Coding for software engineeringSoftware Coding for software engineering
Software Coding for software engineeringssuserb3a23b
 
Exploring Selenium_Appium Frameworks for Seamless Integration with HeadSpin.pdf
Exploring Selenium_Appium Frameworks for Seamless Integration with HeadSpin.pdfExploring Selenium_Appium Frameworks for Seamless Integration with HeadSpin.pdf
Exploring Selenium_Appium Frameworks for Seamless Integration with HeadSpin.pdfkalichargn70th171
 
Comparing Linux OS Image Update Models - EOSS 2024.pdf
Comparing Linux OS Image Update Models - EOSS 2024.pdfComparing Linux OS Image Update Models - EOSS 2024.pdf
Comparing Linux OS Image Update Models - EOSS 2024.pdfDrew Moseley
 
Alfresco TTL#157 - Troubleshooting Made Easy: Deciphering Alfresco mTLS Confi...
Alfresco TTL#157 - Troubleshooting Made Easy: Deciphering Alfresco mTLS Confi...Alfresco TTL#157 - Troubleshooting Made Easy: Deciphering Alfresco mTLS Confi...
Alfresco TTL#157 - Troubleshooting Made Easy: Deciphering Alfresco mTLS Confi...Angel Borroy López
 
What is Fashion PLM and Why Do You Need It
What is Fashion PLM and Why Do You Need ItWhat is Fashion PLM and Why Do You Need It
What is Fashion PLM and Why Do You Need ItWave PLM
 
CRM Contender Series: HubSpot vs. Salesforce
CRM Contender Series: HubSpot vs. SalesforceCRM Contender Series: HubSpot vs. Salesforce
CRM Contender Series: HubSpot vs. SalesforceBrainSell Technologies
 
A healthy diet for your Java application Devoxx France.pdf
A healthy diet for your Java application Devoxx France.pdfA healthy diet for your Java application Devoxx France.pdf
A healthy diet for your Java application Devoxx France.pdfMarharyta Nedzelska
 
Dealing with Cultural Dispersion — Stefano Lambiase — ICSE-SEIS 2024
Dealing with Cultural Dispersion — Stefano Lambiase — ICSE-SEIS 2024Dealing with Cultural Dispersion — Stefano Lambiase — ICSE-SEIS 2024
Dealing with Cultural Dispersion — Stefano Lambiase — ICSE-SEIS 2024StefanoLambiase
 

Dernier (20)

SpotFlow: Tracking Method Calls and States at Runtime
SpotFlow: Tracking Method Calls and States at RuntimeSpotFlow: Tracking Method Calls and States at Runtime
SpotFlow: Tracking Method Calls and States at Runtime
 
Unveiling the Future: Sylius 2.0 New Features
Unveiling the Future: Sylius 2.0 New FeaturesUnveiling the Future: Sylius 2.0 New Features
Unveiling the Future: Sylius 2.0 New Features
 
SuccessFactors 1H 2024 Release - Sneak-Peek by Deloitte Germany
SuccessFactors 1H 2024 Release - Sneak-Peek by Deloitte GermanySuccessFactors 1H 2024 Release - Sneak-Peek by Deloitte Germany
SuccessFactors 1H 2024 Release - Sneak-Peek by Deloitte Germany
 
Machine Learning Software Engineering Patterns and Their Engineering
Machine Learning Software Engineering Patterns and Their EngineeringMachine Learning Software Engineering Patterns and Their Engineering
Machine Learning Software Engineering Patterns and Their Engineering
 
MYjobs Presentation Django-based project
MYjobs Presentation Django-based projectMYjobs Presentation Django-based project
MYjobs Presentation Django-based project
 
Ahmed Motair CV April 2024 (Senior SW Developer)
Ahmed Motair CV April 2024 (Senior SW Developer)Ahmed Motair CV April 2024 (Senior SW Developer)
Ahmed Motair CV April 2024 (Senior SW Developer)
 
Hot Sexy call girls in Patel Nagar🔝 9953056974 🔝 escort Service
Hot Sexy call girls in Patel Nagar🔝 9953056974 🔝 escort ServiceHot Sexy call girls in Patel Nagar🔝 9953056974 🔝 escort Service
Hot Sexy call girls in Patel Nagar🔝 9953056974 🔝 escort Service
 
Odoo Development Company in India | Devintelle Consulting Service
Odoo Development Company in India | Devintelle Consulting ServiceOdoo Development Company in India | Devintelle Consulting Service
Odoo Development Company in India | Devintelle Consulting Service
 
KnowAPIs-UnknownPerf-jaxMainz-2024 (1).pptx
KnowAPIs-UnknownPerf-jaxMainz-2024 (1).pptxKnowAPIs-UnknownPerf-jaxMainz-2024 (1).pptx
KnowAPIs-UnknownPerf-jaxMainz-2024 (1).pptx
 
Intelligent Home Wi-Fi Solutions | ThinkPalm
Intelligent Home Wi-Fi Solutions | ThinkPalmIntelligent Home Wi-Fi Solutions | ThinkPalm
Intelligent Home Wi-Fi Solutions | ThinkPalm
 
SensoDat: Simulation-based Sensor Dataset of Self-driving Cars
SensoDat: Simulation-based Sensor Dataset of Self-driving CarsSensoDat: Simulation-based Sensor Dataset of Self-driving Cars
SensoDat: Simulation-based Sensor Dataset of Self-driving Cars
 
Taming Distributed Systems: Key Insights from Wix's Large-Scale Experience - ...
Taming Distributed Systems: Key Insights from Wix's Large-Scale Experience - ...Taming Distributed Systems: Key Insights from Wix's Large-Scale Experience - ...
Taming Distributed Systems: Key Insights from Wix's Large-Scale Experience - ...
 
Software Coding for software engineering
Software Coding for software engineeringSoftware Coding for software engineering
Software Coding for software engineering
 
Exploring Selenium_Appium Frameworks for Seamless Integration with HeadSpin.pdf
Exploring Selenium_Appium Frameworks for Seamless Integration with HeadSpin.pdfExploring Selenium_Appium Frameworks for Seamless Integration with HeadSpin.pdf
Exploring Selenium_Appium Frameworks for Seamless Integration with HeadSpin.pdf
 
Comparing Linux OS Image Update Models - EOSS 2024.pdf
Comparing Linux OS Image Update Models - EOSS 2024.pdfComparing Linux OS Image Update Models - EOSS 2024.pdf
Comparing Linux OS Image Update Models - EOSS 2024.pdf
 
Alfresco TTL#157 - Troubleshooting Made Easy: Deciphering Alfresco mTLS Confi...
Alfresco TTL#157 - Troubleshooting Made Easy: Deciphering Alfresco mTLS Confi...Alfresco TTL#157 - Troubleshooting Made Easy: Deciphering Alfresco mTLS Confi...
Alfresco TTL#157 - Troubleshooting Made Easy: Deciphering Alfresco mTLS Confi...
 
What is Fashion PLM and Why Do You Need It
What is Fashion PLM and Why Do You Need ItWhat is Fashion PLM and Why Do You Need It
What is Fashion PLM and Why Do You Need It
 
CRM Contender Series: HubSpot vs. Salesforce
CRM Contender Series: HubSpot vs. SalesforceCRM Contender Series: HubSpot vs. Salesforce
CRM Contender Series: HubSpot vs. Salesforce
 
A healthy diet for your Java application Devoxx France.pdf
A healthy diet for your Java application Devoxx France.pdfA healthy diet for your Java application Devoxx France.pdf
A healthy diet for your Java application Devoxx France.pdf
 
Dealing with Cultural Dispersion — Stefano Lambiase — ICSE-SEIS 2024
Dealing with Cultural Dispersion — Stefano Lambiase — ICSE-SEIS 2024Dealing with Cultural Dispersion — Stefano Lambiase — ICSE-SEIS 2024
Dealing with Cultural Dispersion — Stefano Lambiase — ICSE-SEIS 2024
 

Scala Data Pipelines for Music Recommendations

  • 1. January 6, 2015 Scala Data Pipelines for Music Recommendations Chris Johnson @MrChrisJohnson
  • 2. Who am I?? •Chris Johnson – Machine Learning guy from NYC – Focused on music recommendations – Formerly a PhD student at UTAustin
  • 3. Spotify in Numbers 3 •Started in 2006, now available in 58 markets •50+ million active users, 15 million paying subscribers •30+ million songs, 20,000 new songs added per day •1.5 billion playlists •1 TB user data logged per day •900 node Hadoop cluster •10,000+ Hadoop jobs run every day
  • 4. 4 Music Recommendations at Spotify •Discover •Radio •Related Artists
  • 5. How can we find good recommendations? 5 •Manual Curation •Manually Tag Attributes •Audio Content •News, Blogs, Text analysis •Collaborative Filtering
  • 8. Why ? 8 Interview Question
  • 9. The Genre Toplist Problem 9 •Assume we have access to daily log data for all plays on Spotify. •Goal: Calculate the top 1k artists on for each genre based on total daily plays {"User": “userA”, "Date": “2015-01-10", "Artist": “Beyonce", "Track": "Halo", "Genres": ["Pop", "R&B", "Soul"]} {"User": “userB”, "Date": “2015-01-10”, "Artist": "Led Zeppelin”, "Track": "Achilles Last Stand", "Genres": ["Rock", "Blues Rock", "Hard Rock"]} ……….
  • 10. Genre Toplists with Python MapReduce 10
  • 11. 11 Scalding is a Scala library that makes it easy to specify Hadoop MapReduce jobs. Scalding is built on top of Cascading, a Java library that abstracts away low-level Hadoop details. Scalding is comparable to Pig, but offers tight integration with Scala, bringing advantages of Scala to your MapReduce jobs. -Twitter
  • 12. Genre Toplists with Scalding 12
  • 13. Why ? 13 •Data pipeline flows naturally follow the functional paradigm •Productivity without sacrificing performance •Active community and ecosystem -Scalding -Summingbird -Algebird -Spark -Breeze •Many data storage solutions integrate well with JVM -Cassandra -HBase -Voldemort -Datomic
  • 15. Genre Toplists with Scalding 15
  • 16. sortWithTake doesn’t fully sort 16 •Uses PriorityQueueMonoid from Algebird library •What is a Monoid?? -Definition: A Set S and a binary operation • : S x S —> S such that 1. Associativity: For all a, b, and c in S the equation
 (a • b) • c = a • (b • c) holds 2. Identity Element: There exists an element e in S such that for every element a in S, the equations e • a = a • e = a hold •Example: The natural numbers N under the addition operation.
 (1 + 2) + 3 = 1 + (2 + 3)
 0 + 1 = 1 + 0 = 1 class PriorityQueueMonoid[K](max : Int)(implicit ord : Ordering[K]) extends Monoid[PriorityQueue[K]]
  • 17. sortWithTake 17 •Uses PriorityQueueMonoid from Algebird •PriorityQueue aggregations form a commutative monoid! 1. Associative:
 PQ1 = [ (Jay Z, 545), (Miles Davis, 272), …]
 PQ2 = [ (Beyonce, 731), (Kurt Vile, 372), …]
 PQ3 = [ (Twin Shadow, 87), … ]
 PQ1 ++ (PQ2 ++ PQ3) = (PQ1 ++ PQ2) ++ PQ3 2.Commutative: 
 PQ1 ++ PQ2 = PQ2 ++ PQ1 3.Identity: 
 PQ1 ++ EmptyPQ = PQ1 class PriorityQueueMonoid[K](max : Int)(implicit ord : Ordering[K]) extends Monoid[PriorityQueue[K]]
  • 18. sortWithTake 18 •Uses PriorityQueueMonoid from Algebird •Ok, great observation… but what’s the point of all this!?? -All monoid aggregations and reduces can begin on the Mapper side and finish on the Reducer side since the order doesn’t matter! -Scalding implicitly takes care of Mapper side combining and custom combiner -Reduces network traffic to reducers class PriorityQueueMonoid[K](max : Int)(implicit ord : Ordering[K]) extends Monoid[PriorityQueue[K]] reduced traffic
  • 20. How do we store track metadata? 20 •Lots of metadata associated with tracks (100+ columns!) -artist, album, record label, genres, audio features, … •Options: 1. Store each track as one long row with many columns -Sending lots of data over network when you only need 1 or 2 columns 2. Store each column as a separate data source -Jobs require costly joins, especially when requiring many columns •Can we do better?..
  • 21. Apache Parquet to the rescue! 21 •Apache Parquet is a columnar storage format available to any project in the Hadoop ecosystem, regardless of the choice of data processing framework, data model or programming language. •Efficiently read a subset of columns without scanning the entire dataset •Row group: A logical horizontal partitioning of the data into rows. There is no physical structure that is guaranteed for a row group. A row group consists of a column chunk for each column in the dataset. •Column chunk: A chunk of the data for a particular column. These live in a particular row group and is guaranteed to be contiguous in the file. •Predicate push-down: Define predicates (<, >, <=, …) to filter out column chunks or even full row groups, evaluated at Hadoop InputFormat layer before Avro conversion
  • 22. Genre Toplists with Scalding + Parquet 22
  • 23. Driven - job visualization and performance analytics 23
  • 24. Luigi - data plumbing since 2012 24 •Workflow management framework developed by Spotify •Python luigi configuration takes care of dependency resolution, job scheduling, fault tolerance, etc. •Support for Hive queries, MapReduce jobs, python snippets, Scalding, Crunch, Spark, and more! •Like Oozie but without all of the messy XML https://github.com/spotify/luigi
  • 27. So…. back to music recommendations! 27 •Manual Curation •Manually Tag Attributes •Audio Content •News, Blogs, Text analysis •Collaborative Filtering
  • 28. Collaborative Filtering 28 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
  • 29. Implicit Matrix Factorization 29 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 •Aggregate all (user, track) streams into a large matrix •Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight •Why?: Once learned, the top recommendations for a user are the top inner products between their latent factor vector in X and the track latent factor vectors in Y. X YUsers Songs • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • = item latent factor vector
  • 30. Alternating Least Squares 30 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 X YUsers Songs • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • = item latent factor vector Fix tracks •Aggregate all (user, track) streams into a large matrix •Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight •Why?: Once learned, the top recommendations for a user are the top inner products between their latent factor vector in X and the track latent factor vectors in Y.
  • 31. 31 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 X YUsers Songs • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • = item latent factor vector Fix tracks Solve for users •Aggregate all (user, track) streams into a large matrix •Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight •Why?: Once learned, the top recommendations for a user are the top inner products between their latent factor vector in X and the track latent factor vectors in Y. Alternating Least Squares
  • 32. 32 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 X YUsers Songs • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • = item latent factor vector Fix users •Aggregate all (user, track) streams into a large matrix •Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight •Why?: Once learned, the top recommendations for a user are the top inner products between their latent factor vector in X and the track latent factor vectors in Y. Alternating Least Squares
  • 33. 33 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 X YUsers Songs • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • = item latent factor vector Fix users Solve for tracks •Aggregate all (user, track) streams into a large matrix •Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight •Why?: Once learned, the top recommendations for a user are the top inner products between their latent factor vector in X and the track latent factor vectors in Y. Alternating Least Squares
  • 34. 34 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 X YUsers Songs • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • = item latent factor vector Fix users Solve for tracks Repeat until convergence… •Aggregate all (user, track) streams into a large matrix •Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight •Why?: Once learned, the top recommendations for a user are the top inner products between their latent factor vector in X and the track latent factor vectors in Y. Alternating Least Squares
  • 35. 35 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 X YUsers Songs • = bias for user • = bias for item • = regularization parameter • = 1 if user streamed track else 0 • • = user latent factor vector • = item latent factor vector Fix users Solve for tracks Repeat until convergence… •Aggregate all (user, track) streams into a large matrix •Goal: Approximate binary preference matrix by the inner product of 2 smaller matrices by minimizing the weighted RMSE (root mean squared error) using a function of total plays as weight •Why?: Once learned, the top recommendations for a user are the top inner products between their latent factor vector in X and the track latent factor vectors in Y. Alternating Least Squares
  • 36. Matrix Factorization with MapReduce 36 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
  • 37. Matrix Factorization with MapReduce 37 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
  • 38. 38 •Fast and general purpose cluster computing system •Provides high-level apis in Java, Scala, and Python •Takes advantage of in-memory caching to reduce I/O bottleneck of Hadoop MapReduce •MLlib: Scalable Machine Learning library packaged with Spark -Collaborative Filtering and Matrix Factorization -Classification and Regression -Clustering -Optimization Primitives •Spark Streaming: Real time, scalable, fault-tolerant stream processing •Spark SQL: allows relational queries expressed in SQL, HiveQL, or Scala to be executed using Spark
  • 39. Matrix Factorization with Spark 39 streams user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 •Partition streams matrix into user (row) and item (column) blocks, partition, and cache -Unlike with the MapReduce implementation, ratings are never shuffled across the network! •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
  • 40. Matrix Factorization with Spark 40 user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 •Partition streams matrix into user (row) and item (column) blocks, partition, and cache -Unlike with the MapReduce implementation, ratings are never shuffled across the network! •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 streams
  • 41. Matrix Factorization with Spark 41 user vectors item vectors •Partition streams matrix into user (row) and item (column) blocks, partition, and cache -Unlike with the MapReduce implementation, ratings are never shuffled across the network! •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 streams
  • 42. Matrix Factorization with Spark 42 user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY •Partition streams matrix into user (row) and item (column) blocks, partition, and cache -Unlike with the MapReduce implementation, ratings are never shuffled across the network! •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 streams
  • 43. Matrix Factorization with Spark 43 user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY •Partition streams matrix into user (row) and item (column) blocks, partition, and cache -Unlike with the MapReduce implementation, ratings are never shuffled across the network! •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 streams
  • 44. Matrix Factorization with Spark 44 user vectors item vectors worker 1 worker 2 worker 3 worker 4 worker 5 worker 6 YtY YtY YtY YtY YtY YtY •Partition streams matrix into user (row) and item (column) blocks, partition, and cache -Unlike with the MapReduce implementation, ratings are never shuffled across the network! •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 streams
  • 46. Scala Breeze 46 •Native Scala numerical processing library •Linear Algebra -Matrix operations -Operator overloading and syntactic sugar •Sampling from Probably Distributions •Numerical Optimization •Plotting and Visualizations •Numpy for Scala
  • 47. Zeppelin + Spark + Parquet for ETL 47
  • 48. Zeppelin + Spark + Parquet for ETL 48
  • 49. What should I be worried about? 49 •Multiple “right” ways to do the same thing •Implicits can make code difficult to navigate •Learning curve can be tough •Avoid flattening before a join •Be aware that Scala default collections are immutable (though mutable versions are also available) •Use monoid reduces and aggregations where possible and avoid folds •Be patient with the compiler