Merge multiple columns in a Spark DataFrame [Java]

Question

How to combine multiple columns (say 3) from a DataFrame in a single column (in a new DataFrame) where each row becomes a Spark DenseVector? Similar to this thread but in Java and with a few tweaks mentioned below.

I tried using a UDF like this:

private UDF3<Double, Double, Double, Row> toColumn = new UDF3<Double, Double, Double, Row>() {

    private static final long serialVersionUID = 1L;

    public Row call(Double first, Double second, Double third) throws Exception {           
        Row row = RowFactory.create(Vectors.dense(first, second, third));

        return row; 
    }
};

And then register the UDF:

sqlContext.udf().register("toColumn", toColumn, dataType);

Where the dataType is:

StructType dataType = DataTypes.createStructType(new StructField[]{
    new StructField("bla", new VectorUDT(), false, Metadata.empty()),
    });

When I call this UDF on a DataFrame with 3 columns and print out the schema of the new DataFrame, I get this:

root |-- features: struct (nullable = true) | |-- bla: vector (nullable = false)

The problem here is that I need a vector to be outside, not within a struct. Something like this:

root
 |-- features: vector (nullable = true)

I don't know how to get this since the register function requires the return type of UDF to be DataType (which, in turn, doesn't provide a VectorType)

score 0 · Accepted Answer · answered Jun 09 '16 at 22:01

You actually nested the vector type into a struct manually by using this data type:

new StructField("bla", new VectorUDT(), false, Metadata.empty()),

If you remove the outer StructField, you will get what you want. Of course, in this case, you need to modify a bit the signature of your function definition. That is, you need to return with the type Vector.

Please see below my concrete example of what I mean in the form of a simple JUnit test.

package sample.spark.test;

import org.apache.spark.api.java.JavaSparkContext;
import org.apache.spark.mllib.linalg.Vector;
import org.apache.spark.mllib.linalg.VectorUDT;
import org.apache.spark.mllib.linalg.Vectors;
import org.apache.spark.sql.DataFrame;
import org.apache.spark.sql.RowFactory;
import org.apache.spark.sql.SQLContext;
import org.apache.spark.sql.api.java.UDF3;
import org.apache.spark.sql.types.DataTypes;
import org.apache.spark.sql.types.Metadata;
import org.apache.spark.sql.types.StructField;
import org.junit.Test;

import java.io.Serializable;
import java.util.Arrays;
import java.util.HashSet;
import java.util.Set;

import static org.junit.Assert.assertEquals;
import static org.junit.Assert.assertTrue;

public class ToVectorTest implements Serializable {
  private static final long serialVersionUID = 2L;

  private UDF3<Double, Double, Double, Vector> toColumn = new UDF3<Double, Double, Double, Vector>() {

    private static final long serialVersionUID = 1L;

    public Vector call(Double first, Double second, Double third) throws Exception {
      return Vectors.dense(first, second, third);
    }
  };

  @Test
  public void testUDF() {
    // context
    final JavaSparkContext sc = new JavaSparkContext("local", "ToVectorTest");
    final SQLContext sqlContext = new SQLContext(sc);

    // test input
    final DataFrame input = sqlContext.createDataFrame(
        sc.parallelize(
            Arrays.asList(
                RowFactory.create(1.0, 2.0, 3.0),
                RowFactory.create(4.0, 5.0, 6.0),
                RowFactory.create(7.0, 8.0, 9.0),
                RowFactory.create(10.0, 11.0, 12.0)
            )),
        DataTypes.createStructType(
            Arrays.asList(
                new StructField("feature1", DataTypes.DoubleType, false, Metadata.empty()),
                new StructField("feature2", DataTypes.DoubleType, false, Metadata.empty()),
                new StructField("feature3", DataTypes.DoubleType, false, Metadata.empty())
            )
        )
    );
    input.registerTempTable("input");

    // expected output
    final Set<Vector> expectedOutput = new HashSet<>(Arrays.asList(
        Vectors.dense(1.0, 2.0, 3.0),
        Vectors.dense(4.0, 5.0, 6.0),
        Vectors.dense(7.0, 8.0, 9.0),
        Vectors.dense(10.0, 11.0, 12.0)
    ));

    // processing
    sqlContext.udf().register("toColumn", toColumn, new VectorUDT());
    final DataFrame outputDF = sqlContext.sql("SELECT toColumn(feature1, feature2, feature3) AS x FROM input");
    final Set<Vector> output = new HashSet<>(outputDF.toJavaRDD().map(r -> r.<Vector>getAs("x")).collect());

    // evaluation
    assertEquals(expectedOutput.size(), output.size());
    for (Vector x : output) {
      assertTrue(expectedOutput.contains(x));
    }

    // show the schema and the content
    System.out.println(outputDF.schema());
    outputDF.show();

    sc.stop();
  }
}

That is exactly what I needed. Somehow, I managed to not consider returning a Vector from the UDF and registering the function with the VectorUDT. Thanks Robert! — Rajko, Jun 10 '16 at 08:17

Merge multiple columns in a Spark DataFrame [Java]

1 Answers1