Skip to content

Latest commit

 

History

History
186 lines (151 loc) · 11.5 KB

README.md

File metadata and controls

186 lines (151 loc) · 11.5 KB

sparkboost

This repository contains a distributed implementation based on Apache Spark of AdaBoost.MH and MP-Boost algorithms. MP-Boost is an improved variant of the well known AdaBoost.MH machine learning algorithm. MP-Boost improves original AdaBoost.MH by building classifiers which allows to obtain remarkably better effectiveness and a very similar computational cost at build/classification time.

The software is open source and released under the terms of the Apache License, Version 2.0.

The software allows to build multi-label multiclass classifiers or binary classifiers using AdaBoost.MH or MP-Boost starting from a dataset available in the LibSvm format. A lot of ready datasets in this format are available here.

Software installation

To use the latest release of this software in your Maven projects, in your project POM add the following:

<repositories>

    <repository>
        <id>sparkboost-mvn-repo</id>
        <url>https://raw.github.com/tizfa/sparkboost/mvn-repo/</url>
        <snapshots>
            <enabled>true</enabled>
            <updatePolicy>always</updatePolicy>
        </snapshots>
    </repository>

</repositories>

then in the dependencies list add

<dependency>
    <groupId>tizfa</groupId>
    <artifactId>sparkboost</artifactId>
    <version>1.0.0</version>
</dependency>

Software usage

Using provided command line tools

Currently the software allows to perform multilabel multiclass classification or binary classification over datasets available on in the LibSvm format. The user at learning and classification time must specify if the problem is or not of binary type (usage of -b flag in the available commands). In multiclass problems, the user should also specify if the labels IDs are 0-based or 1-based, i.e. if the number of valid labels is n then the set of valid IDs is in the range [0, 9] included (0-based) or [1,10] included (1-based). To specify if the labels are 0-based, the user can use the flag -z in the available commands.

Software compilation to use command line tools

If you are interested to use the command line tools available with the software, you need to download the latest release sources available from here and next compile them. To perform this task, you need Maven and a Java 8 compiler installed on your machine. Download a copy of this software repository on your machine on a specific folder, go inside that folder and at the command prompt put the following commands:

mvn clean
mvn -P shading package

This set of commands will build a software bundle containing all the necessary Spark libraries. You can find the software bundle in the target directory of the software package.

Building a MP-Boost classifier

To build a MP-Boost classifier, launch this command from prompt:

spark-submit --class it.tizianofagni.sparkboost.MPBoostLearnerExe --master <sparkMasterAddress> /path/to/sparkboost-1.0.0-bundle.jar
    <inputLibSVMTrainingData> <outputModelFile> <numIterations>

where sparkMasterAddress is the name of Spark master host (or local[*] for executing the process locally on your machine), inputLibSVMTrainingData is the input training data file in LibSVM format, outputModelFile is the output file where the generated classifier will be save, and numIterations is the number of iterations used in the algorithm. You can control the degree of parallelism (as the number of partitions concurrently processed) used during the computation by specifying the parameters -lp <numPartitions> for labels partitioning, -dp <numPartitions> for documents partitioning, and -fp <numPartitions> for documents partitioning. If the user does not specify anything, the default values are set to be the total number of cores available in Spark runtime.

Building an AdaBoost.MH classifier

Similar to previous case, to build a AdaBoost.MH classifier launch this command from prompt:

spark-submit --class it.tizianofagni.sparkboost.AdaBoostMHLearnerExe --master <sparkMasterAddress> /path/to/sparkboost-1.0.0-bundle.jar
    <inputLibSVMTrainingData> <outputModelFile> <numIterations>

where the parameters have the same meaning as used when building a MP-Boost classifier.

Using a classifier

To use an already built classifier over a test dataset (it does not matter if the model has been built with MP-Boost or AdaBoost.MH learner, they share the same forrmat for classification models!), use this command:

spark-submit --class it.tizianofagni.sparkboost.BoostClassifierExe --master <sparkMasterAddress>  /path/to/sparkboost-1.0.0-bundle.jar <testDatasetFile> 
    <boostModelDirectory> <boostClassifierOutputDirectory>

where sparkMasterAddress is the name of Spark master host (or local[*] for executing the process locally on your machine), testDatasetFile is the input test data file in LibSVM format, boostModelDirectory is the directory where the learner has saved its model and boostClassifierOutputDirectory is the directory where the classification output results will be saved. By using the parameter -p you can specify the parallelism degree while performing classification: if this parameter is not specified, by default the parallelism degree will be equals to the number of available cores in the Spark runtime.

IMPORTANT NOTE: Each document in the test dataset will get a document ID corresponding at the original row index of the document in the dataset file.

Use case: RCV1v2

We used MP-Boost to build a multilabel classifer to automatically classify textual documents in RCV1v2, a corpus of newswire stories made available by Reuters, Ltd. The dataset (rcv1v2 (topics; full sets)) is available also in libsvm format at page http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multilabel.html

The main characteristics of rcv1v2 (topics; full sets) are:

  • number of classes: 101
  • number of documents: 23149 / 781265 (testing)
  • number of features: 47236

We used the following files for this experimentation:

  • rcv1_topics_train.svm.bz2 for training (23149 documents)
  • rcv1_topics_test_0.svm.bz2 for testing (199328 documents)

We built a MP-Boost classification model using 500 iterations and using a single multicore machine (AMD Fx-8350 8-cores). The training time to build a classification model for all 101 labels and by specifying a parallelismDegree of 8 has been of 1206 seconds. The classification time has been of 61 seconds to classify all 199328 documents. Here are the main results we have obtained in this specific configuration: Precision: 0.833, Recall: 0.699, F1:0.760

Using library API to build your own programs

An example of using the API is given by the provided command line tools. Just watch the source code of classes AdaBoostMHLearnerExe.java, MPBoostLearnerExe.java and BoostClassifierExe.java. Briefly, to build a classifier you can use a code like this:

JavaSparkContext sc = ... // Spark context to use;

// Create and configure AdaBoost.MH learner. For MP-Boost, just use the class
// MPBoostLearner.
AdaBoostMHLearner learner = new AdaBoostMHLearner(sc);
learner.setNumIterations(numIterations);
learner.setNumDocumentsPartitions(numDocsPartitions);
learner.setNumFeaturesPartitions(numFeaturesPartitions);
learner.setNumLabelsPartitions(numLabelsPartitions);

// Build a new classifier. Here we assume that the training data is available in
// the input file which is written in LibSvm format.
BoostClassifier classifier = learner.buildModel(inputFile, labels0Based, binaryProblem);

// Save classifier in outputModelPath using the any valid syntax allowed by Spark/Hadoop.
DataUtils.saveModel(sc, classifier, outputModelPath);

Alternatively, you can build a new classifier by specifying directly the set of training documents to use:

JavaSparkContext sc = ... // Spark context to use;

// Create and configure AdaBoost.MH learner. For MP-Boost, just use the class
// MPBoostLearner.
AdaBoostMHLearner learner = new AdaBoostMHLearner(sc);
// Customize learner (numIterations, num docs partitions, etc.) as in the previous snippet of code.....


// You can prepare yourself the training data by generating an RDD with items
// of type MultilabelPoint.
JavaRDD<MultilabelPoint> trainingData = ...

// Build a new classifier.
BoostClassifier classifier = learner.buildModel(trainingData);

// Save classifier in outputModelPath using any valid syntax allowed by Spark/Hadoop.
DataUtils.saveModel(sc, classifier, outputModelPath);

To load a saved classifier and use it for classification tasks, use the following code:

JavaSparkContext sc = ... // Spark context to use;

// Load boosting classifier from disk.
BoostClassifier classifier = DataUtils.loadModel(sc, inputModel);

// Classify documents contained in "inputFile", a file in libsvm format.
ClassificationResults results = classifier.classifyLibSvmWithResults(sc, inputFile, parallelismDegree, labels0Based, binaryProblem);

// or classify documents available in already defined RDD.
JavaRDD<MultilabelPoint> rdd = ...
results = classifier.classifyWithResults(sc, rdd, parallelismDegree);

// Print results in a StringBuilder.
StringBuilder sb = new StringBuilder();
sb.append("**** Effectiveness\n");
sb.append(results.getCt().toString() + "\n");
sb.append("********\n");
for (int i = 0; i < results.getNumDocs(); i++) {
    int docID = results.getDocuments()[i];
    int[] labels = results.getLabels()[i];
    int[] goldLabels = results.getGoldLabels()[i];
    sb.append("DocID: " + docID + ", Labels assigned: " + Arrays.toString(labels) + ", Labels scores: " + Arrays.toString(results.getScores()[i]) + ", Gold labels: " + Arrays.toString(goldLabels) + "\n");
}

The code above assumes that the classification results are all stored on RAM memory of Spark driver process. In situations where this is not feasible, you can use a code like the following to save classification results directly on secondary storage (e.g. hdfs):

JavaSparkContext sc = ... // Spark context to use;

// Load boosting classifier from disk.
BoostClassifier classifier = DataUtils.loadModel(sc, inputModel);

// Get the docs from somewhere (e.g. LibSVM file)
JavaRDD<MultilabelPoint> docs = ...

// Classify documents and save results to an RDD.
JavaRDD<DocClassificationResults> results = classifier.classify(sc, docs, parallelismDegree);

// Save the RDD on secondary storage.
DataUtils.saveHadoopClassificationResults(outputDir, results);

Software compilation for latest snapshot

If you are interested in using the latest snapshot of the software, you need to have Maven and a Java 8 compiler installed on your machine. Download a copy of this software available in 'develop' branch onto your machine, put it on a specific folder, go inside that folder and at the command prompt put the following commands:

mvn clean
mvn -P devel package

This set of commands will build a software bundle containing all the necessary Spark libraries. You can find the software bundle in the target directory of the software package.