swap training and validation order

mllib/src/main/scala/org/apache/spark/mllib/util/MLUtils.scala

@@ -179,8 +179,8 @@ object MLUtils {
 
   /**
    * Return a k element array of pairs of RDDs with the first element of each pair
-   * containing the validation data, a unique 1/Kth of the data and the second
-   * element, the training data, contain the complement of that.
+   * containing the training data, a complement of the validation data and the second
+   * element, the validation data, containing a unique 1/kth of the data. Where k=numFolds.
    */
   def kFold[T: ClassTag](rdd: RDD[T], numFolds: Int, seed: Int): Array[(RDD[T], RDD[T])] = {
     val numFoldsF = numFolds.toFloat
@@ -189,7 +189,7 @@ object MLUtils {
         complement = false)
       val validation = new PartitionwiseSampledRDD(rdd, sampler, seed)
       val training = new PartitionwiseSampledRDD(rdd, sampler.cloneComplement(), seed)
-      (validation, training)
+      (training, validation)
     }.toArray
   }
 

mllib/src/test/scala/org/apache/spark/mllib/util/MLUtilsSuite.scala

@@ -120,7 +120,7 @@ class MLUtilsSuite extends FunSuite with LocalSparkContext {
       for (seed <- 1 to 5) {
         val foldedRdds = MLUtils.kFold(data, folds, seed)
         assert(foldedRdds.size === folds)
-        foldedRdds.map { case (validation, training) =>
+        foldedRdds.map { case (training, validation) =>
           val result = validation.union(training).collect().sorted
           val validationSize = validation.collect().size.toFloat
           assert(validationSize > 0, "empty validation data")