Scalable and Flexible Machine Learning With ScalaBay Area Scala Enthusiasts MeetupMarch 11, 2013 LinkedIn

2

Who are we?

@BigDataSc @ccsevers

3

Stuff you will see today …

Different types of data scientists – Comparison of different approaches to develop machine learning flows

Code The five tool tool – Why Scala (and its ecosystem) is the best tool

to develop machine learning flows (Hint: MapReduce is functional) Some more code Machine Learning examples – Real life (well … almost) examples

of different machine learning problems Even more code

4

“Good data scientists understand, in a deep way, that the heavy lifting of cleanup and

preparation is not something that gets in the way of solving the problem – it is the

problem!”DJ Patil – Founding member of the LinkedIn data science team

5

The data funnel

Real data is an awful, terrible mess Cleaning often is a process of operating on data, excluding some

data, bucketing data and calculating aggregates about the data

These blocks form the basis of most data flows

Generate map, flatMap, for

Exclude filter

Bucket group, groupBy, groupWith

Aggregate sum, reduce, foldLeft

6

There are many ways to develop data flows

7

The mixer

8

The Mixer Word Count

#wordcount.py

from org.apache.pig.scripting import *

@outputSchema("b: bag{ w: chararray}")

def tokenize(words):

return words.split(" ")

script = """

A = load './input.txt';

B = foreach A generate flatten(tokenize((chararray)$0)) as word;

C = group B by word;

D = foreach C generate group, COUNT(B);

store D into './wordcount’ using AvroStorage("schema");

"""

Pig.compile(script).bind().runSingle()

{"schema": {

"type": "record",

"name": "WordCount",

"fields": [

{

"name": "word",

"type": "string"

},

{

"name": "count",

"type": "int"

}]}}

9

The Mixer Data Scientist

Too many occurrences of code inside strings Three different languages inside a single file User Defined Functions (UDFs) vs. Language Support Not real Python, but Jython (which missing some libraries) This is just a simple word count!

10

The Mixer Data Scientist

Pig is great at extract, transform, load (ETL) … as long as you want to use a function that is already part of the

included library … or you get someone else to write it for you (hello, DataFu!) Realistically you will need to maintain a Pig code base and a code

base in some language which can run on the JVM Pig Latin is a bit funky, missing a lot of core programming language

features Pig Latin is interpreted so you get (limited) type and syntax

checking only at runtime

11

The Expert

12

The Expert Word Count

hadoop fs –get input.txt input.txtcp /mnt/hadoop/input.txt ~/MyProjects/WordCount/input.txt

##!/usr/bin/perluse strict;use warnings;

my %count_of;while (my $line = <>) { #read from file or STDIN  foreach my $word (split /\s+/, $line) {     $count_of{$word}++;  }}print "All words and their counts: \n";for my $word (sort keys %count_of) {  print "'$word': $count_of{$word}\n";}__END__

13

The Scalable Expert – Hadoop Streaming

Lets you use any language you want. Same issues as Java MapReduce with regards to multiple passes,

complicated joins, etc. Always reading from stdin and writing to stdout. Easy to test out on local data

– cat myfile.txt | mymapper.sh | sort | myreducer.sh Actual data may not be as nice. No type checking on input or

output can will lead to problems. The main reason to do this is so you can use a nice interpreted

language to do your processing.

14

The craftsman

15

The Craftsman Word Count

package org.myorg;

import java.io.IOException;

import java.util.*;

import org.apache.hadoop.fs.Path;

import org.apache.hadoop.conf.*;

import org.apache.hadoop.io.*;

import org.apache.hadoop.mapreduce.*;

import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;

import org.apache.hadoop.mapreduce.lib.input.TextInputFormat;

import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;

import org.apache.hadoop.mapreduce.lib.output.TextOutputFormat;

public class WordCount {

public static class Map extends Mapper<LongWritable, Text, Text, IntWritable> {

private final static IntWritable one = new IntWritable(1);

private Text word = new Text();

public void map(LongWritable key, Text value, Context context) throws IOException, InterruptedException {

String line = value.toString();

StringTokenizer tokenizer = new StringTokenizer(line);

while (tokenizer.hasMoreTokens()) {

word.set(tokenizer.nextToken());

context.write(word, one);

}

}

}

public static class Reduce extends Reducer<Text, IntWritable, Text, IntWritable> {

public void reduce(Text key, Iterable<IntWritable> values, Context context)

throws IOException, InterruptedException {

int sum = 0;

for (IntWritable val : values) {

sum += val.get();

}

context.write(key, new IntWritable(sum));

}

}

public static void main(String[] args) throws Exception {

Configuration conf = new Configuration();

Job job = new Job(conf, "wordcount");

job.setOutputKeyClass(Text.class);

job.setOutputValueClass(IntWritable.class);

job.setMapperClass(Map.class);

job.setReducerClass(Reduce.class);

job.setInputFormatClass(TextInputFormat.class);

job.setOutputFormatClass(TextOutputFormat.class);

FileInputFormat.addInputPath(job, new Path(args[0]));

FileOutputFormat.setOutputPath(job, new Path(args[1]));

job.waitForCompletion(true);

}

16

The Craftsman Data Scientist

If you like Java it works fine … until you want to do more than one pass, a complicated join or

anything fancy. Cascading solves many of these problems for you but it is still very

verbose

17

We need a better tool

A five tool tool!

http://en.wikipedia.org/wiki/Willie_Mayshttp://en.wikipedia.org/wiki/Five-tool_player

18

The Pragmatic Data Scientist

Agile – Iterates quickly Productive - Uses the right tool for the right job Correct - Tests as much as he can before the job is even submitted Scalable – Can handle real world problems Simple - Single language to represent Operations, UDFs and Data

19

The Pragmatic Data Scientist

Agile – Iterates quickly Productive - Uses the right tool for the right job Correct - Tests as much as he can before the job is even submitted Scalable – Can handle real world problems Simple - Single language to represent Operations, UDFs and Data

20

Agility – Data is complex

21

Agility – Try before you buy

scala> 1 to 10

res0: Range.Inclusive = Range(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

scala> 1 until 10

res1: Range = Range(1, 2, 3, 4, 5, 6, 7, 8, 9)

scala> res0.slice(3, 5)

res3: scala.collection.immutable.IndexedSeq[Int] = Vector(4, 5)

scala> res0.groupBy(_ % 2)

res4: Map[Int, IndexedSeq[Int]] =

Map(1 -> Vector(1, 3, 5, 7, 9), 0 -> Vector(2, 4, 6, 8, 10))

22

Productivity – Don't reinvent the wheel

23

Productivity – Have the work done for you

Python Collections Operators map reduce filter sum min/max

Scala Collections Operators foreach map flatMap collect find takeWhile dropWhile filter withFilter filterNot splitAt

span partition groupBy forall exists count fold reduce sum product min/max

24

Correctness – how to keep your sanity

25

Scalability – works on more than your machine

Integrates with Hadoop (more than just streaming) Has the support of scalable libraries Parallel by design – not just for M/R flows

26

Simplicity

Paco Nathan, Evil Mad Scientist, Concurrent Inc., @pacoid, says:– “[Scalding] code is compact, simple to understand”– “nearly 1:1 between elements of conceptual flow diagram and function

calls”– “Cascalog and Scalding DSLs leverage the functional aspects of

MapReduce, helping to limit complexity in process” Scala is a functional tool for a fundamentally functional job

27

Let’s count some wordsHadoop basics

28

Let’s count some words

This is the “Hello, World!” of anything tangentially related to Hadoop.

Let’s try it in Scala first without any Hadoop stuff.

val myLines : Seq[String] = ... // get some stuff val myWords = myLines.flatMap(w => w.split("\\s+")) val myWordsGrouped = myWords.groupBy(identity) val countedWords = myWordsGrouped.mapValues(x=>x.size) Now write out the words somehow

val countedWords = myLines.flatMap(_.split("\\s+"))

.groupBy(identity)

.mapValues(_.size)

29

Let’s count a lot of words

I’ve gone to the trouble of rewriting this example to run in Hadoop. Here it is: val myLines : TypedPipe[String] = TextLine(args("input")) val myWords = myLines.flatMap(w => w.split("\\s+")) val myWordsGrouped = myWords.groupBy(identity) val countedWords = myWordsGrouped.mapValueStream(x =>

Iterator(x.size)) We can make this even better. val countedWords = myWordsGrouped.size countedWords.write(TypedTsv[(String,Long)](output))

30

Something for nothing

Other people have already done the hard work to make the previous example run

The previous example is using Scalding, a Scala library to write (mainly) Hadoop MapReduce jobs.

https://github.com/twitter/scalding It even has its own Twitter account, @scalding Created by:

– Avi Bryant @avibryant– Oscar Boykin @posco – Argyris Zymnis @argyris

Tweet them now and tell them how awesome it is … I’ll wait

31

Side by side comparison of local and Hadoop

val myWords = myLines.flatMap(w =>

w.split("\\s+"))

val myWordsGrouped = myLines.groupBy(identity)

val countedWords = myWordsGrouped.

mapValues(x=>x.size)

val myWords =

myLines.flatMap(w =>

w.split("\\s+"))

val myWordsGrouped = myWords.groupBy(identity)

val countedWords = myWordsGrouped.

size

There are some small differences, mainly due to how the underlying Hadoop process needs to happen.

32

Why does this work?

Scala has support for embedded domain specific languages (DSLs) Scalding includes a couple DSLs for specifying Cascading (and by

extension Hadoop) workflows. Info about Cascading: http://www.cascading.org/ One of the Scalding DSLs, the Typed one, is designed to be very

close to the standard Scala collections API It’s not a perfect mapping due to how Cascading and Hadoop work,

but in general it is very easy to write your code locally, change a couple small bits, and have it run on a Hadoop cluster

Scalding also has a local mode if you want the syntactic sugar without fussing with Hadoop

33

DSLs for everyone!

We’re showing you Scalding in this talk, but there are others that are similar.

– Scoobi: https://github.com/NICTA/scoobi– Scrunch: https://github.com/cloudera/crunch/tree/master/scrunch

All three attempt to make using code to written on Scala collections work (almost) seamlessly in Hadoop.

More on DSLs: http://www.scala-lang.org/node/1403 Some guts:

34

Fields based DSL

From com.twitter.scalding.Dsl/**

* This object has all the implicit functions and values that are used

* to make the scalding DSL.

*

* It's useful to import Dsl._ when you are writing scalding code outside

* of a Job.

*/

object Dsl extends FieldConversions with TupleConversions with GeneratedTupleAdders with java.io.Serializable {

implicit def pipeToRichPipe(pipe : Pipe) : RichPipe = new

RichPipe(pipe)

implicit def richPipeToPipe(rp : RichPipe) : Pipe = rp.pipe

}

}

}

35

Typed DSL

From com.twitter.scalding.TDsl/** implicits for the type-safe DSL

* import TDsl._ to get the implicit conversions from Grouping/CoGrouping to Pipe,

* to get the .toTypedPipe method on standard cascading Pipes.

* to get automatic conversion of Mappable[T] to TypedPipe[T]

*/

object TDsl extends Serializable with GeneratedTupleAdders {

implicit def pipeTExtensions(pipe : Pipe) : PipeTExtensions = new

PipeTExtensions(pipe)

implicit def mappableToTypedPipe[T](mappable : Mappable[T])

(implicit flowDef : FlowDef, mode : Mode, conv :

TupleConverter[T]) : TypedPipe[T] = {

TypedPipe.from(mappable)(flowDef, mode, conv)

}

}

36

We did something fancy in the previous example: val countedWords = myGroupedWords.size val countedWords = myGroupedWords.mapValues(x =>

1L).sum val countedWords = myGroupedWords.mapValues(x =>

1L).reduce(implicit mon: Monoid[Long])((l,r) => mon.plus(l,r))

Scalding uses Algebird extensively to make your life easier. Algebird can also be used outside of Scalding with no trouble. Algebird has your favorite things like monoids, monads, bloom

filters, count-min sketches, hyperloglogs, etc.

Algebird – It’s like algebra and a bird

37

Counting words with some extra information

Sometimes we want to know some information about the contexts that words occurred in. At eBay, this is often the category that a term appeared in.

Let’s count words and calculate the entropy of the category distribution for each word.

– If you’re unfamiliar with this type of entropy just think of it as a measure of how concentrated the distribution is.

– If you really like formulas it is: Σi p(xi) log(pi)

http://en.wikipedia.org/wiki/Entropy_%28information_theory%29

38

More code

case class MyAvroOutput(word: String, count: Long,

entropy: Double) extends AvroRecord

TypedTsv[(String,Int)]

.flatMap{case(line,cat) => line.split("\\s+").map(x =>

(x,Map(cat->1L))}

.group

.sum

.map{ case(word, dist) =>

val total: Double = dist.values.sum

val entropy = (-1)*dist.values.map{ count =>

(count/total)*math.log(count/total)}.sum

MyAvroOutput(word,total.toLong,entropy)

}

.write(PackedAvroSource[MyAvroOutput](output)) Math is great

39

The reason why you are hereMachine Learning Examples

40

How much should we charge for a Titanic insurance?

Classification case study

41

Titanic II case study

We want to sell life insurance to passengers of Titanic II All we have is data from Titanic I We have to be able to explain why we charge the prices we do

(damn regulators!)

http://commons.wikimedia.org

42

Titanic I Data

Cabin class – e.g. 1st, 2nd, 3rd .. Name – String Age – Integer Embark place – String Destination – String Room – Integer Ticket – Integer Gender – Male or Female

43

Titanic Model

http://www.dtreg.com/

44

Classifier code

object Titanic {

def main(args: Array[String]) = {

// parse data

val reader = new CSVReader(new FileReader(

"src/main/data/titanic.csv"))

val passengers = reader.readAll.tail.map(Passenger(_))

val instances = passengers.map(_.getInstance).toSet

// build tree

val treeBuilder = new TreeBuilder

val tree = treeBuilder.buildTree(instances)

// print tree

tree.dump(System.out)

}

}

45

Titanic Model

http://www.dtreg.com/

46

Let’s cluster some eBay keywords.

Clustering case study

47

Motivation

eBay, like any large site, has a massive number of unique queries every day

Identifying groups of queries based on user behavior might help us to understand the individual queries better

For queries we are unsure of we can even try and match them into a cluster that contains queries we know a lot about.

We can use behavioral things like:– number of searches– number of clicks– number of subsequent bids, buys– number of exits– etc

48

Let’s use Mahout

Apache Mahout, http://mahout.apache.org/, @ApacheMahout, is a powerful machine learning and data mining library that works with Hadoop.

It has a ton of great stuff in it, but many of the drawbacks of using Java MapReduce apply.

It uses some proprietary data formats (is your data in VectorWritable SequenceFiles?)

Luckily for us, there are some nice things that work as standalone pieces.

Coming in release 0.8, there is an excellent single pass k-means clustering algorithm we can use.

49

Let’s use Mahout, inside Scalding

lazy val clust = new StreamingKMeans(new FastProjectionSearch(new EuclideanDistanceMeasure,5,10),

args("sloppyclusters").toInt, (10e-6).asInstanceOf[Float])

var count = 0;

val sloppyClusters =

TextLine(args("input"))

.map{ str =>

val vec = str.split("\t").map(_.toDouble)

val cent = new Centroid(count, new DenseVector(vec))

count += 1

cent

}

.toPipe('centroids)

// This won't work with the current build, coming soon though

.unorderedFoldTo[StreamingKMeans,Centroid]('centroids->’clusters)(clust){(cl,cent) =>

cl.cluster(cent); cl}

.toTypedPipe[StreamingKMeans](Dsl.intFields(Seq(0)))

.flatMap(c => c.iterator.asScala.toIterable)

50

Let’s use Mahout, inside Scalding

val finalClusters = sloppyClusters.groupAll

.mapValueStream{centList =>

lazy val bclusterer = new BallKMeans(new BruteSearch(

new EuclideanDistanceMeasure),

args("numclusters").toInt, 100)

bclusterer.cluster(centList.toList.asJava)

bclusterer.iterator.asScala

}

.values

51

Results

These are primarily eBay head queries. Remember that the clustering algorithm knows nothing about the text in the query.

Sample groups:– chanel, tory burch, diamond ring, kathy van zeeland handbags, ...– ipad 4th generation, samsung galaxy s iii, iphone 4 s, nexus 4, ipad

mini, ...– kohls coupons, lowes coupons– jcrew, cole haan, diesel, banana republic, gucci, burberry, brooks

brothers, …– ferrari, utility trailer, polaris ranger, porsche 911, dump truck, bmw m3,

chainsaw, rv, chevelle, vw bus, dodge charger, ...– paypal account, ebay.com, apple touch icon precomposed.png,

paypal, undefined, ps3%2520games, michael%2520kors

52

Clustering Takeaway

There are some excellent libraries that exist, and even fit the functional model

Scala and Scalding will help you work around the rough edges and integrate them into your data flow, rather than having to create new data flows

Being able to prototype locally and in the Scala REPL saves massive amounts of developer time

53

Using LinkedIn endorsement data to rank Scala experts

Matrix API case study

54

LinkedIn Endorsements

55

Page Rank Algorithm

http://commons.wikimedia.org

56

Prepare Datadef prepareData = {

// read endorsements and transform to edges

val ends = readFile[Endorsement]("endorsements")

.filter(_.skill == "Scala")

.map(e => (e.sender, e.recipient, 1))

.write(TSV(”edges"))

}

def getDominantEigenVector = { … } // outputs to “ranks” (memberId, rank)

def getMembers = {

// get Bay Area members

val members = readLatest[Member]("members")

.filter(_.getRegionCode == 84)

.groupBy(_.getMemberId.toLong)

// join ranks and members

readFile[Ranks](”ranks”).withReducers(10).join(members).toTypedPipe

.map{ case (id, ((_, rank), m)) =>

(rank, m.getMemberId, m.getFirstName, m.getLastName, m.getHeadline) }

.groupAll.sortBy(_._1).reverse.values

.write(TextLine("talk/scalaRanks"))

}

57

Matrix API

mat.mapValues( func ): Matrix mat.filterValues( func ) : Matrix mat.getRow( ind ) : RowVector mat.reduceRowVectors{ f } :

RowVector mat.sumRowVectors :

RowVector mat.mapRows{ func } : Matrix mat.topRowElems( k ) : Matrix mat.rowL1Normalize : Matrix mat.rowL2Normalize : Matrix

rowMeanCentering : Matrix rowSizeAveStdev : Matrix matrix1 * matrix2 : Matrix matrix / scalar(Scalar) : Matrix elemWiseOp( mat2 ){ func } mat1.hProd( matrix2 ) : Matrix mat1.zip( mat2/r/c ) : Matrix matrix.nonZerosWith( sclr ) matrix.trace : Scalar matrix.sum : Scalar matrix.transpose : Matrix mat.diagonal : DiagonalMatrix

58

Time for Results!Endorsements Page Rank

59

1.

2.

3.

4.

5.

6.

7.

8.

9.

10.

60

13.

28.

35.

38.

48.

61

SummaryOnly one slide left!

62

Stuff you have seen today …

There are many ways to develop machine learning programs, none of them are perfect

Scala which reflects the 20 years of evolution since Java's invention, and Scalding which is the same for vanilla MapReduce, are a much better alternative

Machine learning is fun and not necessarily complicated

63