Github user smurching commented on a diff in the pull request:
https://github.com/apache/spark/pull/19433#discussion_r151017375
--- Diff:
mllib/src/main/scala/org/apache/spark/ml/tree/impl/SplitUtils.scala ---
@@ -0,0 +1,215 @@
+/*
+ * 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.tree.impl
+
+import org.apache.spark.ml.tree.{CategoricalSplit, Split}
+import org.apache.spark.mllib.tree.impurity.ImpurityCalculator
+import org.apache.spark.mllib.tree.model.ImpurityStats
+
+/** Utility methods for choosing splits during local & distributed tree
training. */
+private[impl] object SplitUtils {
+
+ /** Sorts ordered feature categories by label centroid, returning an
ordered list of categories */
+ private def sortByCentroid(
+ binAggregates: DTStatsAggregator,
+ featureIndex: Int,
+ featureIndexIdx: Int): List[Int] = {
+ /* Each bin is one category (feature value).
+ * The bins are ordered based on centroidForCategories, and this
ordering determines which
+ * splits are considered. (With K categories, we consider K - 1
possible splits.)
+ *
+ * centroidForCategories is a list: (category, centroid)
+ */
+ val numCategories = binAggregates.metadata.numBins(featureIndex)
+ val nodeFeatureOffset = binAggregates.getFeatureOffset(featureIndexIdx)
+
+ val centroidForCategories = Range(0, numCategories).map { featureValue
=>
+ val categoryStats =
+ binAggregates.getImpurityCalculator(nodeFeatureOffset,
featureValue)
+ val centroid = ImpurityUtils.getCentroid(binAggregates.metadata,
categoryStats)
+ (featureValue, centroid)
+ }
+ // TODO(smurching): How to handle logging statements like these?
+ // logDebug("Centroids for categorical variable: " +
centroidForCategories.mkString(","))
+ // bins sorted by centroids
+ val categoriesSortedByCentroid =
centroidForCategories.toList.sortBy(_._2).map(_._1)
+ // logDebug("Sorted centroids for categorical variable = " +
+ // categoriesSortedByCentroid.mkString(","))
+ categoriesSortedByCentroid
+ }
+
+ /**
+ * Find the best split for an unordered categorical feature at a single
node.
+ *
+ * Algorithm:
+ * - Considers all possible subsets (exponentially many)
+ *
+ * @param featureIndex Global index of feature being split.
+ * @param featureIndexIdx Index of feature being split within subset of
features for current node.
+ * @param featureSplits Array of splits for the current feature
+ * @param parentCalculator Optional: ImpurityCalculator containing
impurity stats for current node
+ * @return (best split, statistics for split) If no valid split was
found, the returned
+ * ImpurityStats instance will be invalid (have member valid =
false).
+ */
+ private[impl] def chooseUnorderedCategoricalSplit(
+ binAggregates: DTStatsAggregator,
+ featureIndex: Int,
+ featureIndexIdx: Int,
+ featureSplits: Array[Split],
+ parentCalculator: Option[ImpurityCalculator] = None): (Split,
ImpurityStats) = {
+ // Unordered categorical feature
+ val nodeFeatureOffset = binAggregates.getFeatureOffset(featureIndexIdx)
+ val numSplits = binAggregates.metadata.numSplits(featureIndex)
+ var parentCalc = parentCalculator
+ val (bestFeatureSplitIndex, bestFeatureGainStats) =
+ Range(0, numSplits).map { splitIndex =>
+ val leftChildStats =
binAggregates.getImpurityCalculator(nodeFeatureOffset, splitIndex)
+ val rightChildStats = binAggregates.getParentImpurityCalculator()
+ .subtract(leftChildStats)
+ val gainAndImpurityStats =
ImpurityUtils.calculateImpurityStats(parentCalc,
+ leftChildStats, rightChildStats, binAggregates.metadata)
+ // Compute parent stats once, when considering first split for
current feature
+ if (parentCalc.isEmpty) {
+ parentCalc = Some(gainAndImpurityStats.impurityCalculator)
+ }
+ (splitIndex, gainAndImpurityStats)
+ }.maxBy(_._2.gain)
+ (featureSplits(bestFeatureSplitIndex), bestFeatureGainStats)
+
+ }
+
+ /**
+ * Choose splitting rule: feature value <= threshold
+ *
+ * @return (best split, statistics for split) If the best split
actually puts all instances
+ * in one leaf node, then it will be set to None. If no valid
split was found, the
+ * returned ImpurityStats instance will be invalid (have member
valid = false)
+ */
+ private[impl] def chooseContinuousSplit(
+ binAggregates: DTStatsAggregator,
+ featureIndex: Int,
+ featureIndexIdx: Int,
+ featureSplits: Array[Split],
+ parentCalculator: Option[ImpurityCalculator] = None): (Split,
ImpurityStats) = {
+ // For a continuous feature, bins are already sorted for splitting
+ // Number of "categories" = number of bins
+ val sortedCategories = Range(0,
binAggregates.metadata.numBins(featureIndex)).toList
+ // Get & return best split info
+ val (bestFeatureSplitIndex, bestFeatureGainStats) =
orderedSplitHelper(binAggregates,
+ featureIndex, featureIndexIdx, sortedCategories, parentCalculator)
+ (featureSplits(bestFeatureSplitIndex), bestFeatureGainStats)
+ }
+
+ /**
+ * Computes the index of the best split for an ordered feature.
+ * @param parentCalculator Optional: ImpurityCalculator containing
impurity stats for current node
+ */
+ private def orderedSplitHelper(
+ binAggregates: DTStatsAggregator,
+ featureIndex: Int,
+ featureIndexIdx: Int,
+ categoriesSortedByCentroid: List[Int],
+ parentCalculator: Option[ImpurityCalculator]): (Int, ImpurityStats)
= {
+ // Cumulative sum (scanLeft) of bin statistics.
+ // Afterwards, binAggregates for a bin is the sum of aggregates for
+ // that bin + all preceding bins.
+ assert(!binAggregates.metadata.isUnordered(featureIndex))
+ val numSplits = binAggregates.metadata.numSplits(featureIndex)
+ val nodeFeatureOffset = binAggregates.getFeatureOffset(featureIndexIdx)
+ var splitIndex = 0
+ while (splitIndex < numSplits) {
+ val currentCategory = categoriesSortedByCentroid(splitIndex)
+ val nextCategory = categoriesSortedByCentroid(splitIndex + 1)
+ binAggregates.mergeForFeature(nodeFeatureOffset, nextCategory,
currentCategory)
+ splitIndex += 1
+ }
+ // lastCategory = index of bin with total aggregates for this (node,
feature)
+ val lastCategory = categoriesSortedByCentroid.last
+
+ // Find best split.
+ var parentCalc = parentCalculator
+ Range(0, numSplits).map { splitIndex =>
+ val featureValue = categoriesSortedByCentroid(splitIndex)
+ val leftChildStats =
+ binAggregates.getImpurityCalculator(nodeFeatureOffset,
featureValue)
+ val rightChildStats =
--- End diff --
Exactly, it's the parentCalc minus the left child stats. Since
`ImpurityCalculator.subtract()` updates the impurity calculator in place, we
call `binAggregates.getParentImpurityCalculator()` to get a copy of the parent
impurity calculator, then subtract the left child stats.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
