4

I'd like to make sure I'm training on a stratified sample of my data.

It seems this is supported by Spark 2.1 and earlier versions via JavaPairRDD.sampleByKey(...) and JavaPairRDD.sampleByKeyExact(...) as explained here.

But: My data is stored in a Dataset<Row>, not a JavaPairRDD. The first column is the label, all others are features (imported from a libsvm-formatted file).

What's the easiest way to get a stratified sample of my dataset instance and at the end have a Dataset<Row> again?

In a way this question is related to Dealing with unbalanced datasets in Spark MLlib.

This possible duplicate does not mention Dataset<Row> at all, neither is it in Java. It does not answer my question.

eliasah
  • 39,588
  • 11
  • 124
  • 154
Hendrik
  • 5,085
  • 24
  • 56

1 Answers1

4

Ok, since the answer of the question here was actually not intended for Java, I have rewritten it in Java.

The reasoning is still the same thought. We are still using sampleByKeyExact. There is no out of the box miracle features for now (spark 2.1.0)

So here you go :

package org.awesomespark.examples;

import org.apache.spark.api.java.JavaPairRDD;
import org.apache.spark.api.java.JavaRDD;
import org.apache.spark.api.java.function.PairFunction;
import org.apache.spark.sql.*;
import scala.Tuple2;

import java.util.Map;

public class StratifiedDatasets {
    public static void main(String[] args) {
        SparkSession spark = SparkSession.builder()
                .appName("Stratified Datasets")
                .getOrCreate();

        Dataset<Row> data = spark.read().format("libsvm").load("sample_libsvm_data.txt");

        JavaPairRDD<Double, Row> rdd = data.toJavaRDD().keyBy(x -> x.getDouble(0));
        Map<Double, Double> fractions = rdd.map(Tuple2::_1)
                .distinct()
                .mapToPair((PairFunction<Double, Double, Double>) (Double x) -> new Tuple2(x, 0.8))
                .collectAsMap();

        JavaRDD<Row> sampledRDD = rdd.sampleByKeyExact(false, fractions, 2L).values();
        Dataset<Row> sampledData = spark.createDataFrame(sampledRDD, data.schema());

        sampledData.show();
        sampledData.printSchema();
    }
}

Now let's package and submit :

$ sbt package
[...]
// [success] Total time: 2 s, completed Jan 16, 2017 1:45:51 PM

$ spark-submit --class org.awesomespark.examples.StratifiedDatasets target/scala-2.10/java-stratified-dataset_2.10-1.0.jar 
[...]

// +-----+--------------------+
// |label|            features|
// +-----+--------------------+
// |  0.0|(692,[127,128,129...|
// |  1.0|(692,[158,159,160...|
// |  1.0|(692,[124,125,126...|
// |  1.0|(692,[152,153,154...|
// |  1.0|(692,[151,152,153...|
// |  0.0|(692,[129,130,131...|
// |  1.0|(692,[99,100,101,...|
// |  0.0|(692,[154,155,156...|
// |  0.0|(692,[127,128,129...|
// |  1.0|(692,[154,155,156...|
// |  0.0|(692,[151,152,153...|
// |  1.0|(692,[129,130,131...|
// |  0.0|(692,[154,155,156...|
// |  1.0|(692,[150,151,152...|
// |  0.0|(692,[124,125,126...|
// |  0.0|(692,[152,153,154...|
// |  1.0|(692,[97,98,99,12...|
// |  1.0|(692,[124,125,126...|
// |  1.0|(692,[156,157,158...|
// |  1.0|(692,[127,128,129...|
// +-----+--------------------+
// only showing top 20 rows

// root
//  |-- label: double (nullable = true)
//  |-- features: vector (nullable = true)

For python users, you can also check my answer Stratified sampling with pyspark.

eliasah
  • 39,588
  • 11
  • 124
  • 154