Mahout Tutorial and Hands-on (version 2015)

Apache Mahout – Tutorial (2015) Cataldo Musto, Ph.D.

Corso di Accesso Intelligente all’Informazione ed Elaborazione del Linguaggio Naturale

Università degli Studi di Bari – Dipartimento di Informatica – A.A. 2014/2015

07/01/2015

1

Outline

• What is Mahout ?

– Overview

• How to use Mahout ?

– Hands-on session

2

Part 1

What is Mahout?

3

What is (a) Mahout?

4

an elephant driver

• Mahout is a Java library

– Implementing Machine Learning techniques

5

What is (a) Mahout?



• Clustering

• Classification

• Recommendation

• Frequent ItemSet

6

What is (a) Mahout?



• Clustering

• Classification

• Recommendation

• Frequent ItemSet (removed)

7

What is (a) Mahout?

What can we do?

• Currently Mahout supports mainly four use cases: – Recommendation - takes users' behavior and tries

to find items users might like.

– Clustering - takes e.g. text documents and groups them into groups of topically related documents.

– Classification - learns from existing categorized documents what documents of a specific category look like and is able to assign unlabelled documents to the (hopefully) correct category.

8

Why Mahout?

• Mahout is not the only ML framework

– Weka

– R (http://www.r-project.org/)

• Why do we prefer Mahout?

(http://www.cs.waikato.ac.nz/ml/weka/)

9

http://www.r-project.org/




http://www.cs.waikato.ac.nz/ml/weka/




Why Mahout?


– Apache License

– Good Community

– Good Documentation

10

Why Mahout?


– Apache License

– Good Community


–Scalable

11

Why Mahout?


– Apache License

– Good Community


–Scalable • Based on Hadoop (not mandatory!)

12

Why do we need a scalable framework?

Big Data!

13

When do we need a scalable framework? (e.g. Recommendation Task)

Over 100m user-preferences connections

14

http://mahout.apache.org/users/recommender/recommender-first-timer-faq.html









Use Cases

15

Use Cases

Recommendation Engine on Foursquare 16

Use Cases

User Interest Modeling on Twitter 17

Use Cases

Pattern Mining on Yahoo! (as anti-spam) 18

Algorithms

• Recommendation

– User-based Collaborative Filtering

– Item-based Collaborative Filtering

– Matrix Factorization-based CF

• Several factorization techniques

19

Algorithms

• Clustering

– K-Means

– Fuzzy K-Means

– Streaming K-Means

– etc.

• Topic Modeling

– LDA (Latent Dirichlet Allocation)

20

Algorithms

• Classification

– Logistic Regression

– Bayes (only on Hadoop, from release 0.9)

– Random Forests (only on Hadoop, from release 0.9)

– Hidden Markov Models

– Perceptrons

21

Algorithms

• Other

– Text processing

• Creation of sparse vectors from text

– Dimensionality Reduction techniques

• Principal Component Analysis (PCA)

• Singular Value Decomposition (SVD)

– (and much more.)

22

Mahout in the Apache Software Foundation

23


Original Mahout Project

24


Taste: collaborative filtering framework

25


Lucene: information retrieval software library

26


Hadoop: framework for distributed storage and programming based on MapReduce

27

Next Releases: Watch out!

Hadoop is going to be replaced by Apache Spark

28

X

http://spark.apache.org



General Architecture

Three-tiers architecture (Application, Algorithms and Shared Libraries)

29


Business Logic

30


Data Storage and Shared Libraries

31


External Applications invoking Mahout APIs

32

In this tutorial we will focus on

Recommendation

33

Recommendation

• Mahout implements a Collaborative Filtering framework – Uses historical data (ratings, clicks, and purchases) to

provide recommendations • User-based: recommend items by finding similar users. This is

often harder to scale because of the dynamic nature of users; • Item-based: calculate similarity between items and make

recommendations. Items usually don't change much, so this often can be computed offline; – Popularized by Amazon and others

• Matrix factorization-based: split the original user-item matrix in smaller matrices in order to analyze item rating patterns and learn some latent factors explaining users’ behavior and items characteristics. – Popularized by Netflix Prize

34

Recommendation - Architecture

35

Recommendation Workflow

Inceptive Idea:

A Java/J2EE application invokes a Mahout

Recommender whose DataModel is based on a set of User Preferences that are

built on the ground of a physical DataStore

36

Physical Storage (database, files, etc.)

37



Data Model

38



Data Model

Recommender

39


External Application


Data Model

Recommender

40


Recommendation in Mahout

• Input: raw data (user preferences) • Output: preferences estimation

• Step 1

– Mapping raw data into a DataModel Mahout-compliant

• Step 2 – Tuning recommender components

• Similarity measure, neighborhood, etc.

• Step 3 – Computing rating estimations

• Step 4 – Evaluating recommendation

41

Recommendation Components

• Mahout key abstractions are implemented through Java interfaces : – DataModel interface

• Methods for mapping raw data to a Mahout-compliant form

– UserSimilarity interface • Methods to calculate the degree of correlation between two users

– ItemSimilarity interface • Methods to calculate the degree of correlation between two items

– UserNeighborhood interface • Methods to define the concept of ‘neighborhood’

– Recommender interface • Methods to implement the recommendation step itself

42

Recommendation Components

• Mahout key abstractions are implemented through Java interfaces :

– example: DataModel interface

• Each methods for mapping raw data to a Mahout-compliant form is an implementation of the generic interface

• e.g. MySQLJDBCDataModel feeds a DataModel from a MySQL database

• (and so on)

43

Components: DataModel

• A DataModel is the interface to draw information about user preferences.

• Which sources is it possible to draw? – Database

• MySQLJDBCDataModel, PostgreSQLDataModel • NoSQL databases supported: MongoDBDataModel, CassandraDataModel

– External Files • FileDataModel

– Generic (preferences directly feed through Java code) • GenericDataModel

(They are all implementations of the DataModel interface)

44

• GenericDataModel – Feed through Java calls

• FileDataModel – CSV (Comma Separated Values)

• JDBCDataModel – JDBC Driver – Standard database structure


45

FileDataModel – CSV input

46


• Regardless the source, they all share a common implementation.

• Basic object: Preference

– Preference is a triple (user,item,score)

– Stored in UserPreferenceArray

47


• Basic object: Preference

– Preference is a triple (user,item,score)

– Stored in UserPreferenceArray

• Two implementations

– GenericUserPreferenceArray

• It stores numerical preference, as well.

– BooleanUserPreferenceArray

• It skips numerical preference values.

48

Components: UserSimilarity

• UserSimilarity defines a notion of similarity between two Users. – (respectively) ItemSimilarity defines a notion of similarity

between two Items.

• Which definition of similarity are available?

– Pearson Correlation – Spearman Correlation – Euclidean Distance – Tanimoto Coefficient – LogLikelihood Similarity

– Already implemented!

49

Example: TanimotoDistance

50

Example: CosineSimilarity

51

Different Similarity definitions

influence neighborhood formation

52

Pearson’s vs. Euclidean distance

53


54


55

Components: UserNeighborhood

• Which definition of neighborhood are available? – Nearest N users

• The first N users with the highest similarity are labeled as ‘neighbors’

– Thresholds • Users whose similarity is above a threshold are labeled as ‘neighbors’

– Already implemented!

56

Components: Recommender

• Given a DataModel, a definition of similarity between users (items) and a definition of neighborhood, a recommender produces as output an estimation of relevance for each unseen item

• Which recommendation algorithms are implemented? – User-based CF

– Item-based CF

– SVD-based CF

(and much more…)

57

Recap

• Many implementations of a CF-based recommender! – Different recommendation algorithms – Different neighborhood definitions – Different similarity definitions

• Evaluation fo the different implementations is actually very time-consuming – The strength of Mahout lies in that it is possible to save

time in the evaluation of the different combinations of the parameters!

– Standard interface for the evaluation of a Recommender System

58

Evaluation

• Mahout provides classes for the evaluation of a recommender system

– Prediction-based measures

• Mean Average Error

• RMSE (Root Mean Square Error)

– IR-based measures

• Precision, Recall, F1-measure, F1@n

• NDCG (ranking measure)

59

Evaluation

• Prediction-based Measures

– Class: AverageAbsoluteDifferenceEvaluator

– Method: evaluate()

– Parameters:

• Recommender implementation

• DataModel implementation

• TrainingSet size (e.g. 70%)

• % of the data to use in the evaluation (smaller % for fast prototyping)

60

Evaluation

• IR-based Measures

– Class: GenericRecommenderIRStatsEvaluator

– Method: evaluate()

– Parameters:

• Recommender implementation

• DataModel implementation

• Relevance Threshold (mean+standard deviation)

• % of the data to use in the evaluation (smaller % for fast prototyping)

61

Part 2

How to use Mahout? Hands-on

62

Download Mahout

• Download – The latest Mahout release is 0.9

– Available at: http://archive.apache.org/dist/mahout/0.9/mahout-distribution-0.9.zip

– Extract all the libraries and include them in a new NetBeans (Eclipse) project

• Requirement: Java 1.6.x or greater.

• Hadoop is not mandatory!

63

http://archive.apache.org/dist/mahout/0.9/mahout-distribution-0.9.zip






Important

JavaDoc https://builds.apache.org/job/Mahout-Quality/javadoc/

64

https://builds.apache.org/job/Mahout-Quality/javadoc/



Exercise 1

• Create a Preference object

• Set preferences through some simple Java call

• Print some statistics about preferences

– How many preferences? On which items?

– Wheter a user has expressed preference on a certain item.

– Which one is the item with the highest score?

65

Hints

• Hints about objects to be used:

– Preference

• Methods setUserId, setItemId, setValue;

– GenericUserPreferenceArray

• Dimension = number of preferences to be defined;

• Methods: getIds(), sortByValueReversed(),hasPrefWithItemId(id);

66

Exercise 1: preferences import org.apache.mahout.cf.taste.impl.model.GenericUserPreferenceArray;

import org.apache.mahout.cf.taste.model.Preference;

import org.apache.mahout.cf.taste.model.PreferenceArray;

class CreatePreferenceArray {

private CreatePreferenceArray() {

}

public static void main(String[] args) {

PreferenceArray user1Prefs = new GenericUserPreferenceArray(2);

user1Prefs.setUserID(0, 1L);

user1Prefs.setItemID(0, 101L);

user1Prefs.setValue(0, 2.0f);



Preference pref = user1Prefs.get(1);

System.out.println(pref);

}

} 67

Exercise 1: preferences import org.apache.mahout.cf.taste.impl.model.GenericUserPreferenceArray;

import org.apache.mahout.cf.taste.model.Preference;


class CreatePreferenceArray {

private CreatePreferenceArray() {

}


PreferenceArray user1Prefs = new GenericUserPreferenceArray(2);

user1Prefs.setUserID(0, 1L);





Preference pref = user1Prefs.get(1);

System.out.println(pref);

}

}

Score 2 for Item 101

68

Exercise 2

• Create a DataModel

• Feed the DataModel through some simple Java calls

• Print some statistics about data (how many users, how many items, maximum ratings, etc.)

69

Exercise 2

• Hints about objects to be used: – FastByIdMap

• PreferenceArray stores the preferences of a single user • Where do the preferences of all the users are stored?

– An HashMap? No. – Mahout introduces data structures optimized for

recommendation tasks – HashMap are replaced by FastByIDMap

– Model • Methods: getNumItems(),

getNumUsers(),getMaxPreference() • General statistics about the model.

70

Exercise 2: data model import org.apache.mahout.cf.taste.impl.common.FastByIDMap;

import org.apache.mahout.cf.taste.impl.model.GenericDataModel;

import org.apache.mahout.cf.taste.impl.model.GenericUserPreferenceArray;

import org.apache.mahout.cf.taste.model.DataModel;


class CreateGenericDataModel {

private CreateGenericDataModel() {

}


FastByIDMap<PreferenceArray> preferences = new FastByIDMap<PreferenceArray>();

PreferenceArray prefsForUser1 = new GenericUserPreferenceArray(10);

prefsForUser1.setUserID(0, 1L);

prefsForUser1.setItemID(0, 101L);

prefsForUser1.setValue(0, 3.0f);

prefsForUser1.setItemID(1, 102L);

prefsForUser1.setValue(1, 4.5f);

preferences.put(1L, prefsForUser1);

DataModel model = new GenericDataModel(preferences);

System.out.println(model);

}

}

71

Exercise 3


• Feed the DataModel through a CSV file

• Calculate similarities between users

– CSV file should contain enough data!

72

Exercise 3


– FileDataModel

• Argument: new File with the path of the CSV

– PearsonCorrelationSimilarity, TanimotoCoefficientSimilarity, etc.

• Argument: the model

73

Exercise 3: similarity import org.apache.mahout.cf.taste.impl.similarity.*;

import org.apache.mahout.cf.taste.impl.model.*;

import org.apache.mahout.cf.taste.impl.model.file.FileDatModel;

class Example3_Similarity {

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

// Istanzia il DataModel e crea alcune statistiche

DataModel model = new FileDataModel(new File("intro.csv"));

UserSimilarity pearson = new PearsonCorrelationSimilarity(model);

UserSimilarity euclidean = new EuclideanDistanceSimilarity(model);

System.out.println("Pearson:"+pearson.userSimilarity(1, 2));

System.out.println("Euclidean:"+euclidean.userSimilarity(1, 2));



}

}

74














}

} FileDataModel

75














}

} Similarity Definitions

76














}

} Output

77

Exercise 4





• Generate neighboorhood

• Generate recommendations

78

Exercise 4







– Compare different combinations of parameters!

79

Exercise 4







– Compare different combinations of parameters!

80

Exercise 4


– NearestNUserNeighborhood

– GenericUserBasedRecommender

• Parameters: – data model already shown

– Neighborhood

» Class: NearestNUserNeighborhood(n,similarity,model)

» Class: ThresholdUserNeighborhood(thr,similarity,model)

– similarity measure already shown

81

Exercise 4: First Recommender import org.apache.mahout.cf.taste.impl.model.file.*;

import org.apache.mahout.cf.taste.impl.neighborhood.*;

import org.apache.mahout.cf.taste.impl.recommender.*;

import org.apache.mahout.cf.taste.impl.similarity.*;

import org.apache.mahout.cf.taste.model.*;

import org.apache.mahout.cf.taste.neighborhood.*;

import org.apache.mahout.cf.taste.recommender.*;

import org.apache.mahout.cf.taste.similarity.*;

class RecommenderIntro {

private RecommenderIntro() {

}



UserSimilarity similarity = new PearsonCorrelationSimilarity(model);

UserNeighborhood neighborhood = new NearestNUserNeighborhood(2, similarity, model);

Recommender recommender = new GenericUserBasedRecommender(

model, neighborhood, similarity);

List<RecommendedItem> recommendations = recommender.recommend(1, 1);

for (RecommendedItem recommendation : recommendations) {

System.out.println(recommendation);

}

}

} 82











}










}

}

}

FileDataModel

83











}










}

}

}

2 neighbours

84











}










}

}

}

Top-1 Recommendation for User 1 85

• Download the GroupLens dataset (100k) – Its format is already Mahout compliant

– http://files.grouplens.org/datasets/movielens/ml-100k.zip

• Preparatory Exercise: repeat exercise 3 (similarity calculations) with a bigger dataset

• Next: now we can run the recommendation framework against a state-of-the-art dataset

Exercise 5: MovieLens Recommender

86

http://files.grouplens.org/datasets/movielens/ml-100k.zip




import org.apache.mahout.cf.taste.impl.model.file.*;










}


DataModel model = new FileDataModel(new File("ua.base"));








}

}

}


87

import org.apache.mahout.cf.taste.impl.model.file.*;










}










}

}

}


We can play with parameters! 88


• Analyze Recommender behavior with different combinations of parameters

– Do the recommendations change with a different similarity measure?

– Do the recommendations change with different neighborhood sizes?

– Which one is the best one?

• …. Let’s go to the next exercise!

89

• Evaluate different CF recommender configurations on MovieLens data

• Metrics: RMSE, MAE, Precision

Exercise 6: Recommender Evaluation

90

• Evaluate different CF recommender configurations on MovieLens data

• Metrics: RMSE, MAE

• Hints: useful classes – Implementations of RecommenderEvaluator

interface • AverageAbsoluteDifferenceRecommenderEvaluator

• RMSRecommenderEvaluator


91

• Further Hints: – Use RandomUtils.useTestSeed()to ensure

the consistency among different evaluation runs

– Invoke the evaluate() method

• Parameters

– RecommenderBuilder: recommender instance (as in previous exercises.

– DataModelBuilder: specific criterion for training

– Split Training-Test: double value (e.g. 0.7 for 70%)

– Amount of data to use in the evaluation: double value (e.g 1.0 for 100%)


92

Example 6: evaluation class EvaluatorIntro {

private EvaluatorIntro() {

}


RandomUtils.useTestSeed();


RecommenderEvaluator evaluator = new AverageAbsoluteDifferenceRecommenderEvaluator();

// Build the same recommender for testing that we did last time:

RecommenderBuilder recommenderBuilder = new RecommenderBuilder() {

@Override

public Recommender buildRecommender(DataModel model) throws TasteException {



return new GenericUserBasedRecommender(model, neighborhood, similarity);

}

};

double score = evaluator.evaluate(recommenderBuilder, null, model, 0.7, 1.0);

System.out.println(score);

}

}

Ensures the consistency between different evaluation runs.

93

Exercise 6: evaluation class EvaluatorIntro {


}







@Override





}

};



}

}

94



}







@Override





}

};



}

}

70%training (whole dataset evaluation)

95



}







@Override





}

};



}

}

Recommendation Engine

96

Example 6: evaluation class EvaluatorIntro {


}





RecommenderEvaluator rmse = new RMSEEvaluator();



@Override





}

};



}

} We can add more measures

97



}








@Override





}

};


double rmse = evaluator.evaluate(recommenderBuilder, null, model, 0.7, 1.0);


System.out.println(rmse);

}

} 98

Exercise 7: item-based recommender

• Mahout provides Java classes for building an item-based recommender system

– Amazon-like

– Recommendations are based on similarities among items (generally pre-computed offline)

– Evaluate it with the MovieLens dataset!

99

class IREvaluatorIntro {

private IREvaluatorIntro() {

}








@Override


ItemSimilarity similarity = new PearsonCorrelationSimilarity(model);

return new GenericItemBasedRecommender(model, similarity);

}

};




System.out.println(rmse); }

}

Exercise 7: item-based recommender

100



}








@Override




}

};





} ItemSimilarity

Example 7: item-based recommender

101



}








@Override




}

};





} No Neighborhood definition for item-based recommenders

Example 7: item-based recommender

102

Exercise 8: MF-based recommender

• Mahout provides Java classes for building an a CF recommender system based on state-of-the-art matrix factorization techniques – Class: SVDRecommender

• Parameters: DataModel, Factorizer

• Factorizer: a factorization algorithm – Alternating Least Squares (ALSWRFactorizer)

– SVD++ (SVDPlusPlusFactorizer)

– Stochastic Gradient Descent (ParallelSGDFactorizer)… etc

– Several parameters to tune!

– Evaluate it with the MovieLens dataset!

103



}








@Override


ALSWRFactorizer = new ALSWRFactorizer(model, 10, 0.065, 60);

return new SVDRecommender(model, factorizer);

}

};





}

104




}








@Override


ALSWRFactorizer = new ALSWRFactorizer(model, 10, 0.065, 60);

return new SVDRecommender(model, factorizer);

}

};





}

105


Hyperparameters: latent factors, lambda, iterations

Mahout Strengths

• Fast-prototyping and evaluation

– To evaluate a different configuration of the same algorithm we just need to update a parameter and run again.

– Example

• Different Neighborhood Size

• Different similarity measures, etc.

106

5 minutes to look for the best configuration

107

• Evaluation of CF algorithms through IR measures

• Metrics: Precision, Recall


108

• Evaluation of CF algorithms through IR measures

• Metrics: Precision, Recall

• Hints: useful classes

– GenericRecommenderIRStatsEvaluator

– Evaluate() method

• Same parameters of exercise 6 and 7


109



}




RecommenderIRStatsEvaluator evaluator = new GenericRecommenderIRStatsEvaluator();



@Override





}

};

IRStatistics stats = evaluator.evaluate(recommenderBuilder, null, model, null, 5,

GenericRecommenderIRStatsEvaluator.CHOOSE_THRESHOLD, 1,0);

System.out.println(stats.getPrecision());

System.out.println(stats.getRecall());

System.out.println(stats.getF1());

}

}

Exercise 9: IR-based evaluation

Precision@5 , Recall@5, etc. 110



}







@Override





}

};






}

}


Precision@5 , Recall@5, etc. 111



}







@Override





}

};






}

}


Set Neighborhood to 500 112



}







@Override


UserSimilarity similarity = new EuclideanDistanceSimilarity(model);



}

};






}

}


Set Euclidean Distance 113

• Write a class that automatically runs evaluation with different parameters

– e.g. fixed neighborhood sizes from an Array of values

– Print the best scores and the configuration


114

• Find the best configuration for several datasets

– Download datasets from http://mahout.apache.org/users/basics/collections.html

–Write classes to transform input data in a Mahout-compliant form

– Extend exercise 10!

Exercise 11: more datasets!

115

http://mahout.apache.org/users/basics/collections.html

http://mahout.apache.org/users/basics/collections.html

End. Do you want more?

116

Do you want more?

• Recommendation – Deploy of a Mahout-based Web Recommender

– Integration with Hadoop/Spark

– Integration of content-based information

– Custom similarities, Custom recommenders, Re-scoring functions

• Content-based Recommender Systems through Classification Algorithms!

117

Technology

Mahout Tutorial and Hands-on (version 2015)