Github user mengxr commented on a diff in the pull request:
https://github.com/apache/spark/pull/79#discussion_r10360528
--- Diff:
mllib/src/main/scala/org/apache/spark/mllib/tree/DecisionTree.scala ---
@@ -0,0 +1,915 @@
+/*
+ * 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.mllib.tree
+
+import org.apache.spark.SparkContext._
+import org.apache.spark.rdd.RDD
+import org.apache.spark.mllib.tree.model._
+import org.apache.spark.{SparkContext, Logging}
+import org.apache.spark.mllib.regression.LabeledPoint
+import org.apache.spark.mllib.tree.model.Split
+import scala.util.control.Breaks._
+import org.apache.spark.mllib.tree.configuration.Strategy
+import org.apache.spark.mllib.tree.configuration.QuantileStrategy._
+import org.apache.spark.mllib.tree.configuration.FeatureType._
+import org.apache.spark.mllib.tree.configuration.Algo._
+import org.apache.spark.mllib.tree.impurity.Impurity
+
+/**
+A class that implements a decision tree algorithm for classification and
regression.
+It supports both continuous and categorical features.
+
+@param strategy The configuration parameters for the tree algorithm which
specify the type of
+algorithm (classification,
+regression, etc.), feature type (continuous, categorical), depth of the
tree,
+quantile calculation strategy, etc.
+ */
+class DecisionTree private (val strategy : Strategy) extends Serializable
with Logging {
+
+ /**
+ Method to train a decision tree model over an RDD
+
+ @param input RDD of [[org.apache.spark.mllib.regression.LabeledPoint]]
used as training data
+ for DecisionTree
+ @return a DecisionTreeModel that can be used for prediction
+ */
+ def train(input : RDD[LabeledPoint]) : DecisionTreeModel = {
+
+ //Cache input RDD for speedup during multiple passes
+ input.cache()
+
+ val (splits, bins) = DecisionTree.findSplitsBins(input, strategy)
+ logDebug("numSplits = " + bins(0).length)
+ strategy.numBins = bins(0).length
+
+ val maxDepth = strategy.maxDepth
+
+ val maxNumNodes = scala.math.pow(2,maxDepth).toInt - 1
+ val filters = new Array[List[Filter]](maxNumNodes)
+ filters(0) = List()
+ val parentImpurities = new Array[Double](maxNumNodes)
+ //Dummy value for top node (updated during first split calculation)
+ //parentImpurities(0) = Double.MinValue
+ val nodes = new Array[Node](maxNumNodes)
+
+ logDebug("algo = " + strategy.algo)
+
+ breakable {
+ for (level <- 0 until maxDepth){
+
+ logDebug("#####################################")
+ logDebug("level = " + level)
+ logDebug("#####################################")
+
+ //Find best split for all nodes at a level
+ val splitsStatsForLevel = DecisionTree.findBestSplits(input,
parentImpurities, strategy,
+ level, filters,splits,bins)
+
+ for ((nodeSplitStats, index) <-
splitsStatsForLevel.view.zipWithIndex){
+
+ extractNodeInfo(nodeSplitStats, level, index, nodes)
+ extractInfoForLowerLevels(level, index, maxDepth,
nodeSplitStats, parentImpurities,
+ filters)
+ logDebug("final best split = " + nodeSplitStats._1)
+
+ }
+ require(scala.math.pow(2,level)==splitsStatsForLevel.length)
+
+ val allLeaf = splitsStatsForLevel.forall(_._2.gain <= 0 )
+ logDebug("all leaf = " + allLeaf)
+ if (allLeaf) break
+
+ }
+ }
+
+ val topNode = nodes(0)
+ topNode.build(nodes)
+
+ val decisionTreeModel = {
+ return new DecisionTreeModel(topNode, strategy.algo)
+ }
+
+ return decisionTreeModel
+ }
+
+
+ private def extractNodeInfo(
+ nodeSplitStats: (Split, InformationGainStats),
+ level: Int, index: Int,
+ nodes: Array[Node]) {
+
+ val split = nodeSplitStats._1
+ val stats = nodeSplitStats._2
+ val nodeIndex = scala.math.pow(2, level).toInt - 1 + index
+ val isLeaf = (stats.gain <= 0) || (level == strategy.maxDepth - 1)
+ val node = new Node(nodeIndex, stats.predict, isLeaf, Some(split),
None, None, Some(stats))
+ logDebug("Node = " + node)
+ nodes(nodeIndex) = node
+ }
+
+ private def extractInfoForLowerLevels(
+ level: Int,
+ index: Int,
+ maxDepth: Int,
+ nodeSplitStats: (Split, InformationGainStats),
+ parentImpurities: Array[Double],
+ filters: Array[List[Filter]]) {
+
+ for (i <- 0 to 1) {
+
+ val nodeIndex = scala.math.pow(2, level + 1).toInt - 1 + 2 * index +
i
+
+ if (level < maxDepth - 1) {
+
+ val impurity = if (i == 0) {
+ nodeSplitStats._2.leftImpurity
+ } else {
+ nodeSplitStats._2.rightImpurity
+ }
+
+ logDebug("nodeIndex = " + nodeIndex + ", impurity = " + impurity)
+ parentImpurities(nodeIndex) = impurity
+ val childFilter = new Filter(nodeSplitStats._1, if (i == 0) -1
else 1)
+ filters(nodeIndex) = childFilter :: filters((nodeIndex - 1) / 2)
+
+ for (filter <- filters(nodeIndex)) {
+ logDebug("Filter = " + filter)
+ }
+
+ }
+ }
+ }
+}
+
+object DecisionTree extends Serializable with Logging {
+
+ /**
+ Method to train a decision tree model over an RDD
+
+ @param input RDD of [[org.apache.spark.mllib.regression.LabeledPoint]]
used as training data
+ for DecisionTree
+ @param strategy The configuration parameters for the tree algorithm
which specify the type of algorithm
+ (classification, regression, etc.),
feature type (continuous, categorical),
+ depth of the tree, quantile calculation
strategy, etc.
+ @return a DecisionTreeModel that can be used for prediction
+ */
+ def train(input : RDD[LabeledPoint], strategy : Strategy) :
DecisionTreeModel = {
+ new DecisionTree(strategy).train(input : RDD[LabeledPoint])
+ }
+
+ /**
+ Method to train a decision tree model over an RDD
+
+ @param input RDD of [[org.apache.spark.mllib.regression.LabeledPoint]]
used as training data
+ for DecisionTree
+ @param algo classification or regression
+ @param impurity criterion used for information gain calculation
+ @param maxDepth maximum depth of the tree
+ @return a DecisionTreeModel that can be used for prediction
+ */
+ def train(
+ input : RDD[LabeledPoint],
+ algo : Algo,
+ impurity : Impurity,
+ maxDepth : Int
+ ) : DecisionTreeModel = {
+ val strategy = new Strategy(algo,impurity,maxDepth)
+ new DecisionTree(strategy).train(input : RDD[LabeledPoint])
+ }
+
+
+ /**
+ Method to train a decision tree model over an RDD
+
+ @param input RDD of [[org.apache.spark.mllib.regression.LabeledPoint]]
used as training data
+ for DecisionTree
+ @param algo classification or regression
+ @param impurity criterion used for information gain calculation
+ @param maxDepth maximum depth of the tree
+ @param maxBins maximum number of bins used for splitting features
+ @param quantileCalculationStrategy algorithm for calculating quantiles
+ @param categoricalFeaturesInfo A map storing information about the
categorical variables and the number of discrete
+ values they take. For example, an entry
(n -> k) implies the feature n is
+ categorical with k categories 0, 1, 2,
... , k-1. It's important to note that
+ features are zero-indexed.
+ @return a DecisionTreeModel that can be used for prediction
+ */
+ def train(
+ input : RDD[LabeledPoint],
+ algo : Algo,
+ impurity : Impurity,
+ maxDepth : Int,
+ maxBins : Int,
+ quantileCalculationStrategy : QuantileStrategy,
+ categoricalFeaturesInfo : Map[Int,Int]
+ ) : DecisionTreeModel = {
+ val strategy = new
Strategy(algo,impurity,maxDepth,maxBins,quantileCalculationStrategy,categoricalFeaturesInfo)
+ new DecisionTree(strategy).train(input : RDD[LabeledPoint])
+ }
+
+ /**
+ Returns an Array[Split] of optimal splits for all nodes at a given level
+
+ @param input RDD of [[org.apache.spark.mllib.regression.LabeledPoint]]
used as training data
+ for DecisionTree
+ @param parentImpurities Impurities for all parent nodes for the current
level
+ @param strategy [[org.apache.spark.mllib.tree.configuration.Strategy]]
instance containing
+ parameters for construction the
DecisionTree
+ @param level Level of the tree
+ @param filters Filter for all nodes at a given level
+ @param splits possible splits for all features
+ @param bins possible bins for all features
+
+ @return Array[Split] instance for best splits for all nodes at a given
level.
+ */
+ def findBestSplits(
+ input : RDD[LabeledPoint],
+ parentImpurities : Array[Double],
+ strategy: Strategy,
+ level: Int,
+ filters : Array[List[Filter]],
+ splits : Array[Array[Split]],
+ bins : Array[Array[Bin]]) : Array[(Split,
InformationGainStats)] = {
+
+ //Common calculations for multiple nested methods
+ val numNodes = scala.math.pow(2, level).toInt
+ logDebug("numNodes = " + numNodes)
+ //Find the number of features by looking at the first sample
+ val numFeatures = input.take(1)(0).features.length
+ logDebug("numFeatures = " + numFeatures)
+ val numBins = strategy.numBins
+ logDebug("numBins = " + numBins)
+
+ /*Find the filters used before reaching the current code*/
+ def findParentFilters(nodeIndex: Int): List[Filter] = {
+ if (level == 0) {
+ List[Filter]()
+ } else {
+ val nodeFilterIndex = scala.math.pow(2, level).toInt - 1 +
nodeIndex
+ filters(nodeFilterIndex)
+ }
+ }
+
+ /**
+ Find whether the sample is valid input for the current node.
+ In other words, does it pass through all the filters for the current
node.
+ */
+ def isSampleValid(parentFilters: List[Filter], labeledPoint:
LabeledPoint): Boolean = {
+
+ //Leaf
+ if ((level > 0) & (parentFilters.length == 0) ){
+ return false
+ }
+
+ for (filter <- parentFilters) {
+ val features = labeledPoint.features
+ val featureIndex = filter.split.feature
+ val threshold = filter.split.threshold
+ val comparison = filter.comparison
+ val categories = filter.split.categories
+ val isFeatureContinuous = filter.split.featureType == Continuous
+ val feature = features(featureIndex)
+ if (isFeatureContinuous){
+ comparison match {
+ case(-1) => if (feature > threshold) return false
+ case(1) => if (feature <= threshold) return false
+ }
+ } else {
+ val containsFeature = categories.contains(feature)
+ comparison match {
+ case(-1) => if (!containsFeature) return false
+ case(1) => if (containsFeature) return false
+ }
+
+ }
+ }
+ true
+ }
+
+ /**
+ Finds the right bin for the given feature
+ */
+ def findBin(featureIndex: Int, labeledPoint: LabeledPoint,
isFeatureContinuous : Boolean) : Int = {
+
+ if (isFeatureContinuous){
+ for (binIndex <- 0 until strategy.numBins) {
+ val bin = bins(featureIndex)(binIndex)
+ val lowThreshold = bin.lowSplit.threshold
+ val highThreshold = bin.highSplit.threshold
+ val features = labeledPoint.features
+ if ((lowThreshold < features(featureIndex)) & (highThreshold >=
features(featureIndex))) {
+ return binIndex
+ }
+ }
+ throw new UnknownError("no bin was found for continuous variable.")
+ } else {
+
+ for (binIndex <- 0 until strategy.numBins) {
+ val bin = bins(featureIndex)(binIndex)
+ val category = bin.category
+ val features = labeledPoint.features
+ if (category == features(featureIndex)) {
+ return binIndex
+ }
+ }
+ throw new UnknownError("no bin was found for categorical
variable.")
+
+ }
+
+ }
+
+ /**
+ Finds bins for all nodes (and all features) at a given level
+ k features, l nodes (level = log2(l))
+ Storage label, b11, b12, b13, .., bk, b21, b22, .. ,bl1, bl2, .. ,blk
+ Denotes invalid sample for tree by noting bin for feature 1 as -1
+ */
+ def findBinsForLevel(labeledPoint : LabeledPoint) : Array[Double] = {
+
+
+ // calculating bin index and label per feature per node
+ val arr = new Array[Double](1+(numFeatures * numNodes))
+ arr(0) = labeledPoint.label
+ for (nodeIndex <- 0 until numNodes) {
+ val parentFilters = findParentFilters(nodeIndex)
+ //Find out whether the sample qualifies for the particular node
+ val sampleValid = isSampleValid(parentFilters, labeledPoint)
+ val shift = 1 + numFeatures * nodeIndex
+ if (!sampleValid) {
+ //Add to invalid bin index -1
+ breakable {
+ for (featureIndex <- 0 until numFeatures) {
+ arr(shift+featureIndex) = -1
+ //Breaking since marking one bin is sufficient
+ break()
+ }
+ }
+ } else {
+ for (featureIndex <- 0 until numFeatures) {
+ //logDebug("shift+featureIndex =" + (shift+featureIndex))
+ val isFeatureContinous =
strategy.categoricalFeaturesInfo.get(featureIndex).isEmpty
+ arr(shift + featureIndex) = findBin(featureIndex,
labeledPoint,isFeatureContinous)
+ }
+ }
+
+ }
+ arr
+ }
+
+ /**
+ Performs a sequential aggregation over a partition for classification.
+
+ for p bins, k features, l nodes (level = log2(l)) storage is of the
form:
+ b111_left_count,b111_right_count, .... , ..
+ .. bpk1_left_count, bpk1_right_count, .... , ..
+ .. bpkl_left_count, bpkl_right_count
+
+ @param agg Array[Double] storing aggregate calculation of size
+ 2*numSplits*numFeatures*numNodes for classification
+ @param arr Array[Double] of size 1+(numFeatures*numNodes)
+ @return Array[Double] storing aggregate calculation of size
2*numSplits*numFeatures*numNodes
+ for classification
+ */
+ def classificationBinSeqOp(arr: Array[Double], agg: Array[Double]) {
+ for (node <- 0 until numNodes) {
+ val validSignalIndex = 1 + numFeatures * node
+ val isSampleValidForNode = if (arr(validSignalIndex) != -1) true
else false
+ if (isSampleValidForNode) {
+ val label = arr(0)
+ for (featureIndex <- 0 until numFeatures) {
+ val arrShift = 1 + numFeatures * node
+ val aggShift = 2 * numBins * numFeatures * node
+ val arrIndex = arrShift + featureIndex
+ val aggIndex = aggShift + 2 * featureIndex * numBins +
arr(arrIndex).toInt * 2
+ label match {
+ case (0.0) => agg(aggIndex) = agg(aggIndex) + 1
+ case (1.0) => agg(aggIndex + 1) = agg(aggIndex + 1) + 1
+ }
+ }
+ }
+ }
+ }
+
+ /**
+ Performs a sequential aggregation over a partition for regression.
+
+ for p bins, k features, l nodes (level = log2(l)) storage is of the
form:
+ b111_count,b111_sum, b111_sum_squares .... , ..
+ .. bpk1_count, bpk1_sum, bpk1_sum_squares, .... , ..
+ .. bpkl_count, bpkl_sum, bpkl_sum_squares
+
+ @param agg Array[Double] storing aggregate calculation of size
+ 3*numSplits*numFeatures*numNodes for classification
+ @param arr Array[Double] of size 1+(numFeatures*numNodes)
+ @return Array[Double] storing aggregate calculation of size
3*numSplits*numFeatures*numNodes
+ for regression
+ */
+ def regressionBinSeqOp(arr: Array[Double], agg: Array[Double]) {
+ for (node <- 0 until numNodes) {
+ val validSignalIndex = 1 + numFeatures * node
+ val isSampleValidForNode = if (arr(validSignalIndex) != -1) true
else false
+ if (isSampleValidForNode) {
+ val label = arr(0)
+ for (feature <- 0 until numFeatures) {
+ val arrShift = 1 + numFeatures * node
+ val aggShift = 3 * numBins * numFeatures * node
+ val arrIndex = arrShift + feature
+ val aggIndex = aggShift + 3 * feature * numBins +
arr(arrIndex).toInt * 3
+ //count, sum, sum^2
+ agg(aggIndex) = agg(aggIndex) + 1
+ agg(aggIndex + 1) = agg(aggIndex + 1) + label
+ agg(aggIndex + 2) = agg(aggIndex + 2) + label*label
+ }
+ }
+ }
+ }
+
+ /**
+ Performs a sequential aggregation over a partition.
+
+ for p bins, k features, l nodes (level = log2(l)) storage is of the
form:
+ b111_left_count,b111_right_count, .... , ..
+ .. bpk1_left_count, bpk1_right_count, .... , ..
+ .. bpkl_left_count, bpkl_right_count
+
+ @param agg Array[Double] storing aggregate calculation of size
2*numSplits*numFeatures*numNodes for classification
+ and 3*numSplits*numFeatures*numNodes for
regression
+ @param arr Array[Double] of size 1+(numFeatures*numNodes)
+ @return Array[Double] storing aggregate calculation of size
2*numSplits*numFeatures*numNodes for classification
+ and 3*numSplits*numFeatures*numNodes for regression
+ */
+ def binSeqOp(agg : Array[Double], arr: Array[Double]) : Array[Double]
= {
+ strategy.algo match {
+ case Classification => classificationBinSeqOp(arr, agg)
+ case Regression => regressionBinSeqOp(arr, agg)
+ }
+ agg
+ }
+
+ val binAggregateLength = strategy.algo match {
+ case Classification => 2*numBins * numFeatures * numNodes
+ case Regression => 3*numBins * numFeatures * numNodes
+ }
+ logDebug("binAggregateLength = " + binAggregateLength)
+
+ /**
+ Combines the aggregates from partitions
+ @param agg1 Array containing aggregates from one or more partitions
+ @param agg2 Array containing aggregates from one or more partitions
+
+ @return Combined aggregate from agg1 and agg2
+ */
+ def binCombOp(agg1 : Array[Double], agg2: Array[Double]) :
Array[Double] = {
+ strategy.algo match {
+ case Classification => {
+ val combinedAggregate = new Array[Double](binAggregateLength)
+ for (index <- 0 until binAggregateLength){
+ combinedAggregate(index) = agg1(index) + agg2(index)
+ }
+ combinedAggregate
+ }
+ case Regression => {
+ val combinedAggregate = new Array[Double](binAggregateLength)
+ for (index <- 0 until binAggregateLength){
+ combinedAggregate(index) = agg1(index) + agg2(index)
+ }
+ combinedAggregate
+ }
+ }
+ }
+
+ logDebug("input = " + input.count)
+ val binMappedRDD = input.map(x => findBinsForLevel(x))
+ logDebug("binMappedRDD.count = " + binMappedRDD.count)
+ //calculate bin aggregates
+
+ val binAggregates = {
+
binMappedRDD.aggregate(Array.fill[Double](binAggregateLength)(0))(binSeqOp,binCombOp)
+ }
+ logDebug("binAggregates.length = " + binAggregates.length)
+ //binAggregates.foreach(x => logDebug(x))
+
+ /**
+ * Calculates the information gain for all splits
+ * @param leftNodeAgg left node aggregates
+ * @param featureIndex feature index
+ * @param splitIndex split index
+ * @param rightNodeAgg right node aggregate
+ * @param topImpurity impurity of the parent node
+ * @return information gain and statistics for all splits
+ */
+ def calculateGainForSplit(leftNodeAgg: Array[Array[Double]],
+ featureIndex: Int,
+ splitIndex: Int,
+ rightNodeAgg: Array[Array[Double]],
+ topImpurity: Double) : InformationGainStats
= {
+ strategy.algo match {
+ case Classification => {
+
+ val left0Count = leftNodeAgg(featureIndex)(2 * splitIndex)
+ val left1Count = leftNodeAgg(featureIndex)(2 * splitIndex + 1)
+ val leftCount = left0Count + left1Count
+
+ val right0Count = rightNodeAgg(featureIndex)(2 * splitIndex)
+ val right1Count = rightNodeAgg(featureIndex)(2 * splitIndex + 1)
+ val rightCount = right0Count + right1Count
+
+ val impurity = {
+ if (level > 0) {
+ topImpurity
+ } else {
+ strategy.impurity.calculate(left0Count + right0Count,
left1Count + right1Count)
+ }
+ }
+
+ if (leftCount == 0) {
+ return new
InformationGainStats(0,topImpurity,Double.MinValue,topImpurity,1)
+ }
+ if (rightCount == 0) {
+ return new
InformationGainStats(0,topImpurity,topImpurity,Double.MinValue,0)
+ }
+
+ val leftImpurity = strategy.impurity.calculate(left0Count,
left1Count)
+ val rightImpurity = strategy.impurity.calculate(right0Count,
right1Count)
+
+ val leftWeight = leftCount.toDouble / (leftCount + rightCount)
+ val rightWeight = rightCount.toDouble / (leftCount + rightCount)
+
+ val gain = {
+ if (level > 0) {
+ impurity - leftWeight * leftImpurity - rightWeight *
rightImpurity
+ } else {
+ impurity - leftWeight * leftImpurity - rightWeight *
rightImpurity
+ }
+ }
+
+ //val predict = leftCount / (leftCount + rightCount)
+ val predict = (left1Count + right1Count) / (leftCount +
rightCount)
+
+ new
InformationGainStats(gain,impurity,leftImpurity,rightImpurity,predict)
--- End diff --
put a space between arguments
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
