How to convert spark DataFrame to RDD mllib LabeledPoints?

Question

I tried to apply PCA to my data and then apply RandomForest to the transformed data. However, PCA.transform(data) gave me a DataFrame but I need a mllib LabeledPoints to feed my RandomForest. How can I do that? My code:

    import org.apache.spark.mllib.util.MLUtils
    import org.apache.spark.{SparkConf, SparkContext}
    import org.apache.spark.mllib.tree.RandomForest
    import org.apache.spark.mllib.tree.model.RandomForestModel
    import org.apache.spark.ml.feature.PCA
    import org.apache.spark.mllib.regression.LabeledPoint
    import org.apache.spark.mllib.linalg.Vectors


    val dataset = MLUtils.loadLibSVMFile(sc, "data/mnist/mnist.bz2")

    val splits = dataset.randomSplit(Array(0.7, 0.3))

    val (trainingData, testData) = (splits(0), splits(1))

    val trainingDf = trainingData.toDF()

    val pca = new PCA()
    .setInputCol("features")
    .setOutputCol("pcaFeatures")
    .setK(100)
    .fit(trainingDf)

    val pcaTrainingData = pca.transform(trainingDf)

    val numClasses = 10
    val categoricalFeaturesInfo = Map[Int, Int]()
    val numTrees = 10 // Use more in practice.
    val featureSubsetStrategy = "auto" // Let the algorithm choose.
    val impurity = "gini"
    val maxDepth = 20
    val maxBins = 32

    val model = RandomForest.trainClassifier(pcaTrainingData, numClasses, categoricalFeaturesInfo,
        numTrees, featureSubsetStrategy, impurity, maxDepth, maxBins)


     error: type mismatch;
     found   : org.apache.spark.sql.DataFrame
     required: org.apache.spark.rdd.RDD[org.apache.spark.mllib.regression.LabeledPoint]

I tried the following two possible solutions but they didn't work:

 scala> val pcaTrainingData = trainingData.map(p => p.copy(features = pca.transform(p.features)))
 <console>:39: error: overloaded method value transform with alternatives:
   (dataset: org.apache.spark.sql.DataFrame)org.apache.spark.sql.DataFrame <and>
   (dataset: org.apache.spark.sql.DataFrame,paramMap: org.apache.spark.ml.param.ParamMap)org.apache.spark.sql.DataFrame <and>
   (dataset: org.apache.spark.sql.DataFrame,firstParamPair: org.apache.spark.ml.param.ParamPair[_],otherParamPairs: org.apache.spark.ml.param.ParamPair[_]*)org.apache.spark.sql.DataFrame
  cannot be applied to (org.apache.spark.mllib.linalg.Vector)

And:

     val labeled = pca
    .transform(trainingDf)
    .map(row => LabeledPoint(row.getDouble(0), row(4).asInstanceOf[Vector[Int]]))

     error: type mismatch;
     found   : scala.collection.immutable.Vector[Int]
     required: org.apache.spark.mllib.linalg.Vector

(I have imported org.apache.spark.mllib.linalg.Vectors in the above case)

Any help?

Answer 1

The correct approach here is the second one you tried - mapping each Row into a LabeledPoint to get an RDD[LabeledPoint]. However, it has two mistakes:

1. The correct Vector class (org.apache.spark.mllib.linalg.Vector) does NOT take type arguments (e.g. Vector[Int]) - so even though you had the right import, the compiler concluded that you meant scala.collection.immutable.Vector which DOES.

2. The DataFrame returned from PCA.fit() has 3 columns, and you tried to extract column number 4. For example, showing first 4 lines:

   +-----+--------------------+--------------------+
   |label|            features|         pcaFeatures|
   +-----+--------------------+--------------------+
   |  5.0|(780,[152,153,154...|[880.071111851977...|
   |  1.0|(780,[158,159,160...|[-41.473039034112...|
   |  2.0|(780,[155,156,157...|[931.444898405036...|
   |  1.0|(780,[124,125,126...|[25.5114585648411...|
   +-----+--------------------+--------------------+
   

   To make this easier - I prefer using the column names instead of their indices.

So here's the transformation you need:

val labeled = pca.transform(trainingDf).rdd.map(row => LabeledPoint(
   row.getAs[Double]("label"),   
   row.getAs[org.apache.spark.mllib.linalg.Vector]("pcaFeatures")
))