[SPARK-9698] [ML] Add RInteraction transformer for supporting R-style feature interactions

This is a pre-req for supporting the ":" operator in the RFormula feature transformer.

Design doc from umbrella task: https://docs.google.com/document/d/10NZNSEurN2EdWM31uFYsgayIPfCFHiuIu3pCWrUmP_c/edit

mengxr

Author: Eric Liang <[email protected]>

Closes #7987 from ericl/interaction.

Author: ericl

mllib/src/main/scala/org/apache/spark/ml/feature/Interaction.scala

@@ -0,0 +1,278 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.ml.feature
+
+import scala.collection.mutable.ArrayBuilder
+
+import org.apache.spark.SparkException
+import org.apache.spark.annotation.Experimental
+import org.apache.spark.ml.attribute._
+import org.apache.spark.ml.param._
+import org.apache.spark.ml.param.shared._
+import org.apache.spark.ml.util.Identifiable
+import org.apache.spark.ml.Transformer
+import org.apache.spark.mllib.linalg.{Vector, VectorUDT, Vectors}
+import org.apache.spark.sql.{DataFrame, Row}
+import org.apache.spark.sql.functions._
+import org.apache.spark.sql.types._
+
+/**
+ * :: Experimental ::
+ * Implements the feature interaction transform. This transformer takes in Double and Vector type
+ * columns and outputs a flattened vector of their feature interactions. To handle interaction,
+ * we first one-hot encode any nominal features. Then, a vector of the feature cross-products is
+ * produced.
+ *
+ * For example, given the input feature values `Double(2)` and `Vector(3, 4)`, the output would be
+ * `Vector(6, 8)` if all input features were numeric. If the first feature was instead nominal
+ * with four categories, the output would then be `Vector(0, 0, 0, 0, 3, 4, 0, 0)`.
+ */
+@Experimental
+class Interaction(override val uid: String) extends Transformer
+  with HasInputCols with HasOutputCol {
+
+  def this() = this(Identifiable.randomUID("interaction"))
+
+  /** @group setParam */
+  def setInputCols(values: Array[String]): this.type = set(inputCols, values)
+
+  /** @group setParam */
+  def setOutputCol(value: String): this.type = set(outputCol, value)
+
+  // optimistic schema; does not contain any ML attributes
+  override def transformSchema(schema: StructType): StructType = {
+    validateParams()
+    StructType(schema.fields :+ StructField($(outputCol), new VectorUDT, false))
+  }
+
+  override def transform(dataset: DataFrame): DataFrame = {
+    validateParams()
+    val inputFeatures = $(inputCols).map(c => dataset.schema(c))
+    val featureEncoders = getFeatureEncoders(inputFeatures)
+    val featureAttrs = getFeatureAttrs(inputFeatures)
+
+    def interactFunc = udf { row: Row =>
+      var indices = ArrayBuilder.make[Int]
+      var values = ArrayBuilder.make[Double]
+      var size = 1
+      indices += 0
+      values += 1.0
+      var featureIndex = row.length - 1
+      while (featureIndex >= 0) {
+        val prevIndices = indices.result()
+        val prevValues = values.result()
+        val prevSize = size
+        val currentEncoder = featureEncoders(featureIndex)
+        indices = ArrayBuilder.make[Int]
+        values = ArrayBuilder.make[Double]
+        size *= currentEncoder.outputSize
+        currentEncoder.foreachNonzeroOutput(row(featureIndex), (i, a) => {
+          var j = 0
+          while (j < prevIndices.length) {
+            indices += prevIndices(j) + i * prevSize
+            values += prevValues(j) * a
+            j += 1
+          }
+        })
+        featureIndex -= 1
+      }
+      Vectors.sparse(size, indices.result(), values.result()).compressed
+    }
+
+    val featureCols = inputFeatures.map { f =>
+      f.dataType match {
+        case DoubleType => dataset(f.name)
+        case _: VectorUDT => dataset(f.name)
+        case _: NumericType | BooleanType => dataset(f.name).cast(DoubleType)
+      }
+    }
+    dataset.select(
+      col("*"),
+      interactFunc(struct(featureCols: _*)).as($(outputCol), featureAttrs.toMetadata()))
+  }
+
+  /**
+   * Creates a feature encoder for each input column, which supports efficient iteration over
+   * one-hot encoded feature values. See also the class-level comment of [[FeatureEncoder]].
+   *
+   * @param features The input feature columns to create encoders for.
+   */
+  private def getFeatureEncoders(features: Seq[StructField]): Array[FeatureEncoder] = {
+    def getNumFeatures(attr: Attribute): Int = {
+      attr match {
+        case nominal: NominalAttribute =>
+          math.max(1, nominal.getNumValues.getOrElse(
+            throw new SparkException("Nominal features must have attr numValues defined.")))
+        case _ =>
+          1  // numeric feature
+      }
+    }
+    features.map { f =>
+      val numFeatures = f.dataType match {
+        case _: NumericType | BooleanType =>
+          Array(getNumFeatures(Attribute.fromStructField(f)))
+        case _: VectorUDT =>
+          val attrs = AttributeGroup.fromStructField(f).attributes.getOrElse(
+            throw new SparkException("Vector attributes must be defined for interaction."))
+          attrs.map(getNumFeatures).toArray
+      }
+      new FeatureEncoder(numFeatures)
+    }.toArray
+  }
+
+  /**
+   * Generates ML attributes for the output vector of all feature interactions. We make a best
+   * effort to generate reasonable names for output features, based on the concatenation of the
+   * interacting feature names and values delimited with `_`. When no feature name is specified,
+   * we fall back to using the feature index (e.g. `foo:bar_2_0` may indicate an interaction
+   * between the numeric `foo` feature and a nominal third feature from column `bar`.
+   *
+   * @param features The input feature columns to the Interaction transformer.
+   */
+  private def getFeatureAttrs(features: Seq[StructField]): AttributeGroup = {
+    var featureAttrs: Seq[Attribute] = Nil
+    features.reverse.foreach { f =>
+      val encodedAttrs = f.dataType match {
+        case _: NumericType | BooleanType =>
+          val attr = Attribute.fromStructField(f)
+          encodedFeatureAttrs(Seq(attr), None)
+        case _: VectorUDT =>
+          val group = AttributeGroup.fromStructField(f)
+          encodedFeatureAttrs(group.attributes.get, Some(group.name))
+      }
+      if (featureAttrs.isEmpty) {
+        featureAttrs = encodedAttrs
+      } else {
+        featureAttrs = encodedAttrs.flatMap { head =>
+          featureAttrs.map { tail =>
+            NumericAttribute.defaultAttr.withName(head.name.get + ":" + tail.name.get)
+          }
+        }
+      }
+    }
+    new AttributeGroup($(outputCol), featureAttrs.toArray)
+  }
+
+  /**
+   * Generates the output ML attributes for a single input feature. Each output feature name has
+   * up to three parts: the group name, feature name, and category name (for nominal features),
+   * each separated by an underscore.
+   *
+   * @param inputAttrs The attributes of the input feature.
+   * @param groupName Optional name of the input feature group (for Vector type features).
+   */
+  private def encodedFeatureAttrs(
+      inputAttrs: Seq[Attribute],
+      groupName: Option[String]): Seq[Attribute] = {
+
+    def format(
+        index: Int,
+        attrName: Option[String],
+        categoryName: Option[String]): String = {
+      val parts = Seq(groupName, Some(attrName.getOrElse(index.toString)), categoryName)
+      parts.flatten.mkString("_")
+    }
+
+    inputAttrs.zipWithIndex.flatMap {
+      case (nominal: NominalAttribute, i) =>
+        if (nominal.values.isDefined) {
+          nominal.values.get.map(
+            v => BinaryAttribute.defaultAttr.withName(format(i, nominal.name, Some(v))))
+        } else {
+          Array.tabulate(nominal.getNumValues.get)(
+            j => BinaryAttribute.defaultAttr.withName(format(i, nominal.name, Some(j.toString))))
+        }
+      case (a: Attribute, i) =>
+        Seq(NumericAttribute.defaultAttr.withName(format(i, a.name, None)))
+    }
+  }
+
+  override def copy(extra: ParamMap): Interaction = defaultCopy(extra)
+
+  override def validateParams(): Unit = {
+    require(get(inputCols).isDefined, "Input cols must be defined first.")
+    require(get(outputCol).isDefined, "Output col must be defined first.")
+    require($(inputCols).length > 0, "Input cols must have non-zero length.")
+    require($(inputCols).distinct.length == $(inputCols).length, "Input cols must be distinct.")
+  }
+}
+
+/**
+ * This class performs on-the-fly one-hot encoding of features as you iterate over them. To
+ * indicate which input features should be one-hot encoded, an array of the feature counts
+ * must be passed in ahead of time.
+ *
+ * @param numFeatures Array of feature counts for each input feature. For nominal features this
+ *                    count is equal to the number of categories. For numeric features the count
+ *                    should be set to 1.
+ */
+private[ml] class FeatureEncoder(numFeatures: Array[Int]) {
+  assert(numFeatures.forall(_ > 0), "Features counts must all be positive.")
+
+  /** The size of the output vector. */
+  val outputSize = numFeatures.sum
+
+  /** Precomputed offsets for the location of each output feature. */
+  private val outputOffsets = {
+    val arr = new Array[Int](numFeatures.length)
+    var i = 1
+    while (i < arr.length) {
+      arr(i) = arr(i - 1) + numFeatures(i - 1)
+      i += 1
+    }
+    arr
+  }
+
+  /**
+   * Given an input row of features, invokes the specific function for every non-zero output.
+   *
+   * @param value The row value to encode, either a Double or Vector.
+   * @param f The callback to invoke on each non-zero (index, value) output pair.
+   */
+  def foreachNonzeroOutput(value: Any, f: (Int, Double) => Unit): Unit = value match {
+    case d: Double =>
+      assert(numFeatures.length == 1, "DoubleType columns should only contain one feature.")
+      val numOutputCols = numFeatures.head
+      if (numOutputCols > 1) {
+        assert(
+          d >= 0.0 && d == d.toInt && d < numOutputCols,
+          s"Values from column must be indices, but got $d.")
+        f(d.toInt, 1.0)
+      } else {
+        f(0, d)
+      }
+    case vec: Vector =>
+      assert(numFeatures.length == vec.size,
+        s"Vector column size was ${vec.size}, expected ${numFeatures.length}")
+      vec.foreachActive { (i, v) =>
+        val numOutputCols = numFeatures(i)
+        if (numOutputCols > 1) {
+          assert(
+            v >= 0.0 && v == v.toInt && v < numOutputCols,
+            s"Values from column must be indices, but got $v.")
+          f(outputOffsets(i) + v.toInt, 1.0)
+        } else {
+          f(outputOffsets(i), v)
+        }
+      }
+    case null =>
+      throw new SparkException("Values to interact cannot be null.")
+    case o =>
+      throw new SparkException(s"$o of type ${o.getClass.getName} is not supported.")
+  }
+}