diff options
author | Joseph K. Bradley <joseph.kurata.bradley@gmail.com> | 2014-09-08 09:47:13 -0700 |
---|---|---|
committer | Xiangrui Meng <meng@databricks.com> | 2014-09-08 09:47:13 -0700 |
commit | 711356b422c66e2a80377a9f43fce97282460520 (patch) | |
tree | 203459a199620778ec407845d77a5040767ea2f5 /mllib/src/main | |
parent | 0d1cc4ae42e1f73538dd8b9b1880ca9e5b124108 (diff) | |
download | spark-711356b422c66e2a80377a9f43fce97282460520.tar.gz spark-711356b422c66e2a80377a9f43fce97282460520.tar.bz2 spark-711356b422c66e2a80377a9f43fce97282460520.zip |
[SPARK-3086] [SPARK-3043] [SPARK-3156] [mllib] DecisionTree aggregation improvements
Summary:
1. Variable numBins for each feature [SPARK-3043]
2. Reduced data reshaping in aggregation [SPARK-3043]
3. Choose ordering for ordered categorical features adaptively [SPARK-3156]
4. Changed nodes to use 1-indexing [SPARK-3086]
5. Small clean-ups
Note: This PR looks bigger than it is since I moved several functions from inside findBestSplitsPerGroup to outside of it (to make it clear what was being serialized in the aggregation).
Speedups: This update helps most when many features use few bins but a few features use many bins. Some example results on speedups with 2M examples, 3.5K features (15-worker EC2 cluster):
* Example where old code was reasonably efficient (1/2 continuous, 1/4 binary, 1/4 20-category): 164.813 --> 116.491 sec
* Example where old code wasted many bins (1/10 continuous, 81/100 binary, 9/100 20-category): 128.701 --> 39.334 sec
Details:
(1) Variable numBins for each feature [SPARK-3043]
DecisionTreeMetadata now computes a variable numBins for each feature. It also tracks numSplits.
(2) Reduced data reshaping in aggregation [SPARK-3043]
Added DTStatsAggregator, a wrapper around the aggregate statistics array for easy but efficient indexing.
* Added ImpurityAggregator and ImpurityCalculator classes, to make DecisionTree code more oblivious to the type of impurity.
* Design note: I originally tried creating Impurity classes which stored data and storing the aggregates in an Array[Array[Array[Impurity]]]. However, this led to significant slowdowns, perhaps because of overhead in creating so many objects.
The aggregate statistics are never reshaped, and cumulative sums are computed in-place.
Updated the layout of aggregation functions. The update simplifies things by (1) dividing features into ordered/unordered (instead of ordered/unordered/continuous) and (2) making use of the DTStatsAggregator for indexing.
For this update, the following functions were refactored:
* updateBinForOrderedFeature
* updateBinForUnorderedFeature
* binaryOrNotCategoricalBinSeqOp
* multiclassWithCategoricalBinSeqOp
* regressionBinSeqOp
The above 5 functions were replaced with:
* orderedBinSeqOp
* someUnorderedBinSeqOp
Other changes:
* calculateGainForSplit now treats all feature types the same way.
* Eliminated extractLeftRightNodeAggregates.
(3) Choose ordering for ordered categorical features adaptively [SPARK-3156]
Updated binsToBestSplit():
* This now computes cumulative sums of stats for ordered features.
* For ordered categorical features, it chooses an ordering for categories. (This uses to be done by findSplitsBins.)
* Uses iterators to shorten code and avoid building an Array[Array[InformationGainStats]].
Side effects:
* In findSplitsBins: A sample of the data is only taken for data with continuous features. It is not needed for data with only categorical features.
* In findSplitsBins: splits and bins are no longer pre-computed for ordered categorical features since they are not needed.
* TreePoint binning is simpler for categorical features.
(4) Changed nodes to use 1-indexing [SPARK-3086]
Nodes used to be indexed from 0. Now they are indexed from 1.
Node indexing functions are now collected in object Node (Node.scala).
(5) Small clean-ups
Eliminated functions extractNodeInfo() and extractInfoForLowerLevels() to reduce duplicate code.
Eliminated InvalidBinIndex since it is no longer used.
CC: mengxr manishamde Please let me know if you have thoughts on this—thanks!
Author: Joseph K. Bradley <joseph.kurata.bradley@gmail.com>
Closes #2125 from jkbradley/dt-opt3alt and squashes the following commits:
42c192a [Joseph K. Bradley] Merge branch 'rfs' into dt-opt3alt
d3cc46b [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt3alt
00e4404 [Joseph K. Bradley] optimization for TreePoint construction (pre-computing featureArity and isUnordered as arrays)
425716c [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into rfs
a2acea5 [Joseph K. Bradley] Small optimizations based on profiling
aa4e4df [Joseph K. Bradley] Updated DTStatsAggregator with bug fix (nodeString should not be multiplied by statsSize)
4651154 [Joseph K. Bradley] Changed numBins semantics for unordered features. * Before: numBins = numSplits = (1 << k - 1) - 1 * Now: numBins = 2 * numSplits = 2 * [(1 << k - 1) - 1] * This also involved changing the semantics of: ** DecisionTreeMetadata.numUnorderedBins()
1e3b1c7 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt3alt
1485fcc [Joseph K. Bradley] Made some DecisionTree methods private.
92f934f [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt3alt
e676da1 [Joseph K. Bradley] Updated documentation for DecisionTree
37ca845 [Joseph K. Bradley] Fixed problem with how DecisionTree handles ordered categorical features.
105f8ab [Joseph K. Bradley] Removed commented-out getEmptyBinAggregates from DecisionTree
062c31d [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt3alt
6d32ccd [Joseph K. Bradley] In DecisionTree.binsToBestSplit, changed loops to iterators to shorten code.
807cd00 [Joseph K. Bradley] Finished DTStatsAggregator, a wrapper around the aggregate statistics for easy but hopefully efficient indexing. Modified old ImpurityAggregator classes and renamed them ImpurityCalculator; added ImpurityAggregator classes which work with DTStatsAggregator but do not store data. Unit tests all succeed.
f2166fd [Joseph K. Bradley] still working on DTStatsAggregator
92f7118 [Joseph K. Bradley] Added partly written DTStatsAggregator
fd8df30 [Joseph K. Bradley] Moved some aggregation helpers outside of findBestSplitsPerGroup
d7c53ee [Joseph K. Bradley] Added more doc for ImpurityAggregator
a40f8f1 [Joseph K. Bradley] Changed nodes to be indexed from 1. Tests work.
95cad7c [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt3
5f94342 [Joseph K. Bradley] Added treeAggregate since not yet merged from master. Moved node indexing functions to Node.
61c4509 [Joseph K. Bradley] Fixed bugs from merge: missing DT timer call, and numBins setting. Cleaned up DT Suite some.
3ba7166 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt3
b314659 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt3
9c83363 [Joseph K. Bradley] partial merge but not done yet
45f7ea7 [Joseph K. Bradley] partial merge, not yet done
5fce635 [Joseph K. Bradley] Merge branch 'dt-opt2' into dt-opt3
26d10dd [Joseph K. Bradley] Removed tree/model/Filter.scala since no longer used. Removed debugging println calls in DecisionTree.scala.
356daba [Joseph K. Bradley] Merge branch 'dt-opt1' into dt-opt2
430d782 [Joseph K. Bradley] Added more debug info on binning error. Added some docs.
d036089 [Joseph K. Bradley] Print timing info to logDebug.
e66f1b1 [Joseph K. Bradley] TreePoint * Updated doc * Made some methods private
8464a6e [Joseph K. Bradley] Moved TimeTracker to tree/impl/ in its own file, and cleaned it up. Removed debugging println calls from DecisionTree. Made TreePoint extend Serialiable
a87e08f [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt1
dd4d3aa [Joseph K. Bradley] Mid-process in bug fix: bug for binary classification with categorical features * Bug: Categorical features were all treated as ordered for binary classification. This is possible but would require the bin ordering to be determined on-the-fly after the aggregation. Currently, the ordering is determined a priori and fixed for all splits. * (Temp) Fix: Treat low-arity categorical features as unordered for binary classification. * Related change: I removed most tests for isMulticlass in the code. I instead test metadata for whether there are unordered features. * Status: The bug may be fixed, but more testing needs to be done.
438a660 [Joseph K. Bradley] removed subsampling for mnist8m from DT
86e217f [Joseph K. Bradley] added cache to DT input
e3c84cc [Joseph K. Bradley] Added stuff fro mnist8m to D T Runner
51ef781 [Joseph K. Bradley] Fixed bug introduced by last commit: Variance impurity calculation was incorrect since counts were swapped accidentally
fd65372 [Joseph K. Bradley] Major changes: * Created ImpurityAggregator classes, rather than old aggregates. * Feature split/bin semantics are based on ordered vs. unordered ** E.g.: numSplits = numBins for all unordered features, and numSplits = numBins - 1 for all ordered features. * numBins can differ for each feature
c1565a5 [Joseph K. Bradley] Small DecisionTree updates: * Simplification: Updated calculateGainForSplit to take aggregates for a single (feature, split) pair. * Internal doc: findAggForOrderedFeatureClassification
b914f3b [Joseph K. Bradley] DecisionTree optimization: eliminated filters + small changes
b2ed1f3 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-opt
0f676e2 [Joseph K. Bradley] Optimizations + Bug fix for DecisionTree
3211f02 [Joseph K. Bradley] Optimizing DecisionTree * Added TreePoint representation to avoid calling findBin multiple times. * (not working yet, but debugging)
f61e9d2 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-timing
bcf874a [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-timing
511ec85 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-timing
a95bc22 [Joseph K. Bradley] timing for DecisionTree internals
Diffstat (limited to 'mllib/src/main')
10 files changed, 1188 insertions, 991 deletions
diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/DecisionTree.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/DecisionTree.scala index 5cdd258f6c..dd766c12d2 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/DecisionTree.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/DecisionTree.scala @@ -28,8 +28,9 @@ import org.apache.spark.mllib.tree.configuration.Strategy import org.apache.spark.mllib.tree.configuration.Algo._ import org.apache.spark.mllib.tree.configuration.FeatureType._ import org.apache.spark.mllib.tree.configuration.QuantileStrategy._ -import org.apache.spark.mllib.tree.impl.{DecisionTreeMetadata, TimeTracker, TreePoint} +import org.apache.spark.mllib.tree.impl._ import org.apache.spark.mllib.tree.impurity.{Impurities, Impurity} +import org.apache.spark.mllib.tree.impurity._ import org.apache.spark.mllib.tree.model._ import org.apache.spark.rdd.RDD import org.apache.spark.storage.StorageLevel @@ -65,36 +66,41 @@ class DecisionTree (private val strategy: Strategy) extends Serializable with Lo val retaggedInput = input.retag(classOf[LabeledPoint]) val metadata = DecisionTreeMetadata.buildMetadata(retaggedInput, strategy) logDebug("algo = " + strategy.algo) + logDebug("maxBins = " + metadata.maxBins) // Find the splits and the corresponding bins (interval between the splits) using a sample // of the input data. timer.start("findSplitsBins") val (splits, bins) = DecisionTree.findSplitsBins(retaggedInput, metadata) - val numBins = bins(0).length timer.stop("findSplitsBins") - logDebug("numBins = " + numBins) + logDebug("numBins: feature: number of bins") + logDebug(Range(0, metadata.numFeatures).map { featureIndex => + s"\t$featureIndex\t${metadata.numBins(featureIndex)}" + }.mkString("\n")) // Bin feature values (TreePoint representation). // Cache input RDD for speedup during multiple passes. val treeInput = TreePoint.convertToTreeRDD(retaggedInput, bins, metadata) .persist(StorageLevel.MEMORY_AND_DISK) - val numFeatures = metadata.numFeatures // depth of the decision tree val maxDepth = strategy.maxDepth - // the max number of nodes possible given the depth of the tree - val maxNumNodes = (2 << maxDepth) - 1 + require(maxDepth <= 30, + s"DecisionTree currently only supports maxDepth <= 30, but was given maxDepth = $maxDepth.") + // Number of nodes to allocate: max number of nodes possible given the depth of the tree, plus 1 + val maxNumNodesPlus1 = Node.startIndexInLevel(maxDepth + 1) // Initialize an array to hold parent impurity calculations for each node. - val parentImpurities = new Array[Double](maxNumNodes) + val parentImpurities = new Array[Double](maxNumNodesPlus1) // dummy value for top node (updated during first split calculation) - val nodes = new Array[Node](maxNumNodes) + val nodes = new Array[Node](maxNumNodesPlus1) // Calculate level for single group construction // Max memory usage for aggregates val maxMemoryUsage = strategy.maxMemoryInMB * 1024 * 1024 logDebug("max memory usage for aggregates = " + maxMemoryUsage + " bytes.") - val numElementsPerNode = DecisionTree.getElementsPerNode(metadata, numBins) + // TODO: Calculate memory usage more precisely. + val numElementsPerNode = DecisionTree.getElementsPerNode(metadata) logDebug("numElementsPerNode = " + numElementsPerNode) val arraySizePerNode = 8 * numElementsPerNode // approx. memory usage for bin aggregate array @@ -124,26 +130,29 @@ class DecisionTree (private val strategy: Strategy) extends Serializable with Lo // Find best split for all nodes at a level. timer.start("findBestSplits") - val splitsStatsForLevel = DecisionTree.findBestSplits(treeInput, parentImpurities, - metadata, level, nodes, splits, bins, maxLevelForSingleGroup, timer) + val splitsStatsForLevel: Array[(Split, InformationGainStats)] = + DecisionTree.findBestSplits(treeInput, parentImpurities, + metadata, level, nodes, splits, bins, maxLevelForSingleGroup, timer) timer.stop("findBestSplits") - val levelNodeIndexOffset = (1 << level) - 1 + val levelNodeIndexOffset = Node.startIndexInLevel(level) for ((nodeSplitStats, index) <- splitsStatsForLevel.view.zipWithIndex) { val nodeIndex = levelNodeIndexOffset + index - val isLeftChild = level != 0 && nodeIndex % 2 == 1 - val parentNodeIndex = if (isLeftChild) { // -1 for root node - (nodeIndex - 1) / 2 - } else { - (nodeIndex - 2) / 2 - } + // Extract info for this node (index) at the current level. timer.start("extractNodeInfo") - extractNodeInfo(nodeSplitStats, level, index, nodes) + val split = nodeSplitStats._1 + val stats = nodeSplitStats._2 + val isLeaf = (stats.gain <= 0) || (level == strategy.maxDepth) + val node = new Node(nodeIndex, stats.predict, isLeaf, Some(split), None, None, Some(stats)) + logDebug("Node = " + node) + nodes(nodeIndex) = node timer.stop("extractNodeInfo") + if (level != 0) { // Set parent. - if (isLeftChild) { + val parentNodeIndex = Node.parentIndex(nodeIndex) + if (Node.isLeftChild(nodeIndex)) { nodes(parentNodeIndex).leftNode = Some(nodes(nodeIndex)) } else { nodes(parentNodeIndex).rightNode = Some(nodes(nodeIndex)) @@ -151,11 +160,21 @@ class DecisionTree (private val strategy: Strategy) extends Serializable with Lo } // Extract info for nodes at the next lower level. timer.start("extractInfoForLowerLevels") - extractInfoForLowerLevels(level, index, maxDepth, nodeSplitStats, parentImpurities) + if (level < maxDepth) { + val leftChildIndex = Node.leftChildIndex(nodeIndex) + val leftImpurity = stats.leftImpurity + logDebug("leftChildIndex = " + leftChildIndex + ", impurity = " + leftImpurity) + parentImpurities(leftChildIndex) = leftImpurity + + val rightChildIndex = Node.rightChildIndex(nodeIndex) + val rightImpurity = stats.rightImpurity + logDebug("rightChildIndex = " + rightChildIndex + ", impurity = " + rightImpurity) + parentImpurities(rightChildIndex) = rightImpurity + } timer.stop("extractInfoForLowerLevels") - logDebug("final best split = " + nodeSplitStats._1) + logDebug("final best split = " + split) } - require((1 << level) == splitsStatsForLevel.length) + require(Node.maxNodesInLevel(level) == splitsStatsForLevel.length) // Check whether all the nodes at the current level at leaves. val allLeaf = splitsStatsForLevel.forall(_._2.gain <= 0) logDebug("all leaf = " + allLeaf) @@ -171,7 +190,7 @@ class DecisionTree (private val strategy: Strategy) extends Serializable with Lo logDebug("#####################################") // Initialize the top or root node of the tree. - val topNode = nodes(0) + val topNode = nodes(1) // Build the full tree using the node info calculated in the level-wise best split calculations. topNode.build(nodes) @@ -183,47 +202,6 @@ class DecisionTree (private val strategy: Strategy) extends Serializable with Lo new DecisionTreeModel(topNode, strategy.algo) } - /** - * Extract the decision tree node information for the given tree level and node index - */ - private def extractNodeInfo( - nodeSplitStats: (Split, InformationGainStats), - level: Int, - index: Int, - nodes: Array[Node]): Unit = { - val split = nodeSplitStats._1 - val stats = nodeSplitStats._2 - val nodeIndex = (1 << level) - 1 + index - val isLeaf = (stats.gain <= 0) || (level == strategy.maxDepth) - val node = new Node(nodeIndex, stats.predict, isLeaf, Some(split), None, None, Some(stats)) - logDebug("Node = " + node) - nodes(nodeIndex) = node - } - - /** - * Extract the decision tree node information for the children of the node - */ - private def extractInfoForLowerLevels( - level: Int, - index: Int, - maxDepth: Int, - nodeSplitStats: (Split, InformationGainStats), - parentImpurities: Array[Double]): Unit = { - - if (level >= maxDepth) { - return - } - - val leftNodeIndex = (2 << level) - 1 + 2 * index - val leftImpurity = nodeSplitStats._2.leftImpurity - logDebug("leftNodeIndex = " + leftNodeIndex + ", impurity = " + leftImpurity) - parentImpurities(leftNodeIndex) = leftImpurity - - val rightNodeIndex = leftNodeIndex + 1 - val rightImpurity = nodeSplitStats._2.rightImpurity - logDebug("rightNodeIndex = " + rightNodeIndex + ", impurity = " + rightImpurity) - parentImpurities(rightNodeIndex) = rightImpurity - } } object DecisionTree extends Serializable with Logging { @@ -425,9 +403,6 @@ object DecisionTree extends Serializable with Logging { impurity, maxDepth, maxBins) } - - private val InvalidBinIndex = -1 - /** * Returns an array of optimal splits for all nodes at a given level. Splits the task into * multiple groups if the level-wise training task could lead to memory overflow. @@ -436,12 +411,12 @@ object DecisionTree extends Serializable with Logging { * @param parentImpurities Impurities for all parent nodes for the current level * @param metadata Learning and dataset metadata * @param level Level of the tree - * @param splits possible splits for all features - * @param bins possible bins for all features + * @param splits possible splits for all features, indexed (numFeatures)(numSplits) + * @param bins possible bins for all features, indexed (numFeatures)(numBins) * @param maxLevelForSingleGroup the deepest level for single-group level-wise computation. * @return array (over nodes) of splits with best split for each node at a given level. */ - protected[tree] def findBestSplits( + private[tree] def findBestSplits( input: RDD[TreePoint], parentImpurities: Array[Double], metadata: DecisionTreeMetadata, @@ -475,14 +450,147 @@ object DecisionTree extends Serializable with Logging { } /** + * Get the node index corresponding to this data point. + * This function mimics prediction, passing an example from the root node down to a node + * at the current level being trained; that node's index is returned. + * + * @param node Node in tree from which to classify the given data point. + * @param binnedFeatures Binned feature vector for data point. + * @param bins possible bins for all features, indexed (numFeatures)(numBins) + * @param unorderedFeatures Set of indices of unordered features. + * @return Leaf index if the data point reaches a leaf. + * Otherwise, last node reachable in tree matching this example. + * Note: This is the global node index, i.e., the index used in the tree. + * This index is different from the index used during training a particular + * set of nodes in a (level, group). + */ + private def predictNodeIndex( + node: Node, + binnedFeatures: Array[Int], + bins: Array[Array[Bin]], + unorderedFeatures: Set[Int]): Int = { + if (node.isLeaf) { + node.id + } else { + val featureIndex = node.split.get.feature + val splitLeft = node.split.get.featureType match { + case Continuous => { + val binIndex = binnedFeatures(featureIndex) + val featureValueUpperBound = bins(featureIndex)(binIndex).highSplit.threshold + // bin binIndex has range (bin.lowSplit.threshold, bin.highSplit.threshold] + // We do not need to check lowSplit since bins are separated by splits. + featureValueUpperBound <= node.split.get.threshold + } + case Categorical => { + val featureValue = binnedFeatures(featureIndex) + node.split.get.categories.contains(featureValue) + } + case _ => throw new RuntimeException(s"predictNodeIndex failed for unknown reason.") + } + if (node.leftNode.isEmpty || node.rightNode.isEmpty) { + // Return index from next layer of nodes to train + if (splitLeft) { + Node.leftChildIndex(node.id) + } else { + Node.rightChildIndex(node.id) + } + } else { + if (splitLeft) { + predictNodeIndex(node.leftNode.get, binnedFeatures, bins, unorderedFeatures) + } else { + predictNodeIndex(node.rightNode.get, binnedFeatures, bins, unorderedFeatures) + } + } + } + } + + /** + * Helper for binSeqOp, for data which can contain a mix of ordered and unordered features. + * + * For ordered features, a single bin is updated. + * For unordered features, bins correspond to subsets of categories; either the left or right bin + * for each subset is updated. + * + * @param agg Array storing aggregate calculation, with a set of sufficient statistics for + * each (node, feature, bin). + * @param treePoint Data point being aggregated. + * @param nodeIndex Node corresponding to treePoint. Indexed from 0 at start of (level, group). + * @param bins possible bins for all features, indexed (numFeatures)(numBins) + * @param unorderedFeatures Set of indices of unordered features. + */ + private def mixedBinSeqOp( + agg: DTStatsAggregator, + treePoint: TreePoint, + nodeIndex: Int, + bins: Array[Array[Bin]], + unorderedFeatures: Set[Int]): Unit = { + // Iterate over all features. + val numFeatures = treePoint.binnedFeatures.size + val nodeOffset = agg.getNodeOffset(nodeIndex) + var featureIndex = 0 + while (featureIndex < numFeatures) { + if (unorderedFeatures.contains(featureIndex)) { + // Unordered feature + val featureValue = treePoint.binnedFeatures(featureIndex) + val (leftNodeFeatureOffset, rightNodeFeatureOffset) = + agg.getLeftRightNodeFeatureOffsets(nodeIndex, featureIndex) + // Update the left or right bin for each split. + val numSplits = agg.numSplits(featureIndex) + var splitIndex = 0 + while (splitIndex < numSplits) { + if (bins(featureIndex)(splitIndex).highSplit.categories.contains(featureValue)) { + agg.nodeFeatureUpdate(leftNodeFeatureOffset, splitIndex, treePoint.label) + } else { + agg.nodeFeatureUpdate(rightNodeFeatureOffset, splitIndex, treePoint.label) + } + splitIndex += 1 + } + } else { + // Ordered feature + val binIndex = treePoint.binnedFeatures(featureIndex) + agg.nodeUpdate(nodeOffset, featureIndex, binIndex, treePoint.label) + } + featureIndex += 1 + } + } + + /** + * Helper for binSeqOp, for regression and for classification with only ordered features. + * + * For each feature, the sufficient statistics of one bin are updated. + * + * @param agg Array storing aggregate calculation, with a set of sufficient statistics for + * each (node, feature, bin). + * @param treePoint Data point being aggregated. + * @param nodeIndex Node corresponding to treePoint. Indexed from 0 at start of (level, group). + * @return agg + */ + private def orderedBinSeqOp( + agg: DTStatsAggregator, + treePoint: TreePoint, + nodeIndex: Int): Unit = { + val label = treePoint.label + val nodeOffset = agg.getNodeOffset(nodeIndex) + // Iterate over all features. + val numFeatures = agg.numFeatures + var featureIndex = 0 + while (featureIndex < numFeatures) { + val binIndex = treePoint.binnedFeatures(featureIndex) + agg.nodeUpdate(nodeOffset, featureIndex, binIndex, label) + featureIndex += 1 + } + } + + /** * Returns an array of optimal splits for a group of nodes at a given level * * @param input Training data: RDD of [[org.apache.spark.mllib.tree.impl.TreePoint]] * @param parentImpurities Impurities for all parent nodes for the current level * @param metadata Learning and dataset metadata * @param level Level of the tree - * @param splits possible splits for all features - * @param bins possible bins for all features, indexed as (numFeatures)(numBins) + * @param nodes Array of all nodes in the tree. Used for matching data points to nodes. + * @param splits possible splits for all features, indexed (numFeatures)(numSplits) + * @param bins possible bins for all features, indexed (numFeatures)(numBins) * @param numGroups total number of node groups at the current level. Default value is set to 1. * @param groupIndex index of the node group being processed. Default value is set to 0. * @return array of splits with best splits for all nodes at a given level. @@ -527,88 +635,22 @@ object DecisionTree extends Serializable with Logging { // numNodes: Number of nodes in this (level of tree, group), // where nodes at deeper (larger) levels may be divided into groups. - val numNodes = (1 << level) / numGroups + val numNodes = Node.maxNodesInLevel(level) / numGroups logDebug("numNodes = " + numNodes) - // Find the number of features by looking at the first sample. - val numFeatures = metadata.numFeatures - logDebug("numFeatures = " + numFeatures) - - // numBins: Number of bins = 1 + number of possible splits - val numBins = bins(0).length - logDebug("numBins = " + numBins) - - val numClasses = metadata.numClasses - logDebug("numClasses = " + numClasses) - - val isMulticlass = metadata.isMulticlass - logDebug("isMulticlass = " + isMulticlass) - - val isMulticlassWithCategoricalFeatures = metadata.isMulticlassWithCategoricalFeatures - logDebug("isMultiClassWithCategoricalFeatures = " + isMulticlassWithCategoricalFeatures) + logDebug("numFeatures = " + metadata.numFeatures) + logDebug("numClasses = " + metadata.numClasses) + logDebug("isMulticlass = " + metadata.isMulticlass) + logDebug("isMulticlassWithCategoricalFeatures = " + + metadata.isMulticlassWithCategoricalFeatures) // shift when more than one group is used at deep tree level val groupShift = numNodes * groupIndex - /** - * Get the node index corresponding to this data point. - * This function mimics prediction, passing an example from the root node down to a node - * at the current level being trained; that node's index is returned. - * - * @return Leaf index if the data point reaches a leaf. - * Otherwise, last node reachable in tree matching this example. - */ - def predictNodeIndex(node: Node, binnedFeatures: Array[Int]): Int = { - if (node.isLeaf) { - node.id - } else { - val featureIndex = node.split.get.feature - val splitLeft = node.split.get.featureType match { - case Continuous => { - val binIndex = binnedFeatures(featureIndex) - val featureValueUpperBound = bins(featureIndex)(binIndex).highSplit.threshold - // bin binIndex has range (bin.lowSplit.threshold, bin.highSplit.threshold] - // We do not need to check lowSplit since bins are separated by splits. - featureValueUpperBound <= node.split.get.threshold - } - case Categorical => { - val featureValue = if (metadata.isUnordered(featureIndex)) { - binnedFeatures(featureIndex) - } else { - val binIndex = binnedFeatures(featureIndex) - bins(featureIndex)(binIndex).category - } - node.split.get.categories.contains(featureValue) - } - case _ => throw new RuntimeException(s"predictNodeIndex failed for unknown reason.") - } - if (node.leftNode.isEmpty || node.rightNode.isEmpty) { - // Return index from next layer of nodes to train - if (splitLeft) { - node.id * 2 + 1 // left - } else { - node.id * 2 + 2 // right - } - } else { - if (splitLeft) { - predictNodeIndex(node.leftNode.get, binnedFeatures) - } else { - predictNodeIndex(node.rightNode.get, binnedFeatures) - } - } - } - } - - def nodeIndexToLevel(idx: Int): Int = { - if (idx == 0) { - 0 - } else { - math.floor(math.log(idx) / math.log(2)).toInt - } - } - - // Used for treePointToNodeIndex - val levelOffset = (1 << level) - 1 + // Used for treePointToNodeIndex to get an index for this (level, group). + // - Node.startIndexInLevel(level) gives the global index offset for nodes at this level. + // - groupShift corrects for groups in this level before the current group. + val globalNodeIndexOffset = Node.startIndexInLevel(level) + groupShift /** * Find the node index for the given example. @@ -619,661 +661,254 @@ object DecisionTree extends Serializable with Logging { if (level == 0) { 0 } else { - val globalNodeIndex = predictNodeIndex(nodes(0), treePoint.binnedFeatures) - // Get index for this (level, group). - globalNodeIndex - levelOffset - groupShift - } - } - - /** - * Increment aggregate in location for (node, feature, bin, label). - * - * @param treePoint Data point being aggregated. - * @param agg Array storing aggregate calculation, of size: - * numClasses * numBins * numFeatures * numNodes. - * Indexed by (node, feature, bin, label) where label is the least significant bit. - * @param nodeIndex Node corresponding to treePoint. Indexed from 0 at start of (level, group). - */ - def updateBinForOrderedFeature( - treePoint: TreePoint, - agg: Array[Double], - nodeIndex: Int, - featureIndex: Int): Unit = { - // Update the left or right count for one bin. - val aggIndex = - numClasses * numBins * numFeatures * nodeIndex + - numClasses * numBins * featureIndex + - numClasses * treePoint.binnedFeatures(featureIndex) + - treePoint.label.toInt - agg(aggIndex) += 1 - } - - /** - * Increment aggregate in location for (nodeIndex, featureIndex, [bins], label), - * where [bins] ranges over all bins. - * Updates left or right side of aggregate depending on split. - * - * @param nodeIndex Node corresponding to treePoint. Indexed from 0 at start of (level, group). - * @param treePoint Data point being aggregated. - * @param agg Indexed by (left/right, node, feature, bin, label) - * where label is the least significant bit. - * The left/right specifier is a 0/1 index indicating left/right child info. - * @param rightChildShift Offset for right side of agg. - */ - def updateBinForUnorderedFeature( - nodeIndex: Int, - featureIndex: Int, - treePoint: TreePoint, - agg: Array[Double], - rightChildShift: Int): Unit = { - val featureValue = treePoint.binnedFeatures(featureIndex) - // Update the left or right count for one bin. - val aggShift = - numClasses * numBins * numFeatures * nodeIndex + - numClasses * numBins * featureIndex + - treePoint.label.toInt - // Find all matching bins and increment their values - val featureCategories = metadata.featureArity(featureIndex) - val numCategoricalBins = (1 << featureCategories - 1) - 1 - var binIndex = 0 - while (binIndex < numCategoricalBins) { - val aggIndex = aggShift + binIndex * numClasses - if (bins(featureIndex)(binIndex).highSplit.categories.contains(featureValue)) { - agg(aggIndex) += 1 - } else { - agg(rightChildShift + aggIndex) += 1 - } - binIndex += 1 - } - } - - /** - * Helper for binSeqOp. - * - * @param agg Array storing aggregate calculation, of size: - * numClasses * numBins * numFeatures * numNodes. - * Indexed by (node, feature, bin, label) where label is the least significant bit. - * @param treePoint Data point being aggregated. - * @param nodeIndex Node corresponding to treePoint. Indexed from 0 at start of (level, group). - */ - def binaryOrNotCategoricalBinSeqOp( - agg: Array[Double], - treePoint: TreePoint, - nodeIndex: Int): Unit = { - // Iterate over all features. - var featureIndex = 0 - while (featureIndex < numFeatures) { - updateBinForOrderedFeature(treePoint, agg, nodeIndex, featureIndex) - featureIndex += 1 - } - } - - val rightChildShift = numClasses * numBins * numFeatures * numNodes - - /** - * Helper for binSeqOp. - * - * @param agg Array storing aggregate calculation. - * For ordered features, this is of size: - * numClasses * numBins * numFeatures * numNodes. - * For unordered features, this is of size: - * 2 * numClasses * numBins * numFeatures * numNodes. - * @param treePoint Data point being aggregated. - * @param nodeIndex Node corresponding to treePoint. Indexed from 0 at start of (level, group). - */ - def multiclassWithCategoricalBinSeqOp( - agg: Array[Double], - treePoint: TreePoint, - nodeIndex: Int): Unit = { - val label = treePoint.label - // Iterate over all features. - var featureIndex = 0 - while (featureIndex < numFeatures) { - if (metadata.isUnordered(featureIndex)) { - updateBinForUnorderedFeature(nodeIndex, featureIndex, treePoint, agg, rightChildShift) - } else { - updateBinForOrderedFeature(treePoint, agg, nodeIndex, featureIndex) - } - featureIndex += 1 - } - } - - /** - * Performs a sequential aggregation over a partition for regression. - * For l nodes, k features, - * the count, sum, sum of squares of one of the p bins is incremented. - * - * @param agg Array storing aggregate calculation, updated by this function. - * Size: 3 * numBins * numFeatures * numNodes - * @param treePoint Data point being aggregated. - * @param nodeIndex Node corresponding to treePoint. Indexed from 0 at start of (level, group). - * @return agg - */ - def regressionBinSeqOp(agg: Array[Double], treePoint: TreePoint, nodeIndex: Int): Unit = { - val label = treePoint.label - // Iterate over all features. - var featureIndex = 0 - while (featureIndex < numFeatures) { - // Update count, sum, and sum^2 for one bin. - val binIndex = treePoint.binnedFeatures(featureIndex) - val aggIndex = - 3 * numBins * numFeatures * nodeIndex + - 3 * numBins * featureIndex + - 3 * binIndex - agg(aggIndex) += 1 - agg(aggIndex + 1) += label - agg(aggIndex + 2) += label * label - featureIndex += 1 + val globalNodeIndex = + predictNodeIndex(nodes(1), treePoint.binnedFeatures, bins, metadata.unorderedFeatures) + globalNodeIndex - globalNodeIndexOffset } } /** * Performs a sequential aggregation over a partition. - * For l nodes, k features, - * For classification: - * Either the left count or the right count of one of the bins is - * incremented based upon whether the feature is classified as 0 or 1. - * For regression: - * The count, sum, sum of squares of one of the bins is incremented. * - * @param agg Array storing aggregate calculation, updated by this function. - * Size for classification: - * numClasses * numBins * numFeatures * numNodes for ordered features, or - * 2 * numClasses * numBins * numFeatures * numNodes for unordered features. - * Size for regression: - * 3 * numBins * numFeatures * numNodes. + * Each data point contributes to one node. For each feature, + * the aggregate sufficient statistics are updated for the relevant bins. + * + * @param agg Array storing aggregate calculation, with a set of sufficient statistics for + * each (node, feature, bin). * @param treePoint Data point being aggregated. * @return agg */ - def binSeqOp(agg: Array[Double], treePoint: TreePoint): Array[Double] = { + def binSeqOp( + agg: DTStatsAggregator, + treePoint: TreePoint): DTStatsAggregator = { val nodeIndex = treePointToNodeIndex(treePoint) // If the example does not reach this level, then nodeIndex < 0. // If the example reaches this level but is handled in a different group, // then either nodeIndex < 0 (previous group) or nodeIndex >= numNodes (later group). if (nodeIndex >= 0 && nodeIndex < numNodes) { - if (metadata.isClassification) { - if (isMulticlassWithCategoricalFeatures) { - multiclassWithCategoricalBinSeqOp(agg, treePoint, nodeIndex) - } else { - binaryOrNotCategoricalBinSeqOp(agg, treePoint, nodeIndex) - } + if (metadata.unorderedFeatures.isEmpty) { + orderedBinSeqOp(agg, treePoint, nodeIndex) } else { - regressionBinSeqOp(agg, treePoint, nodeIndex) + mixedBinSeqOp(agg, treePoint, nodeIndex, bins, metadata.unorderedFeatures) } } agg } - // Calculate bin aggregate length for classification or regression. - val binAggregateLength = numNodes * getElementsPerNode(metadata, numBins) - 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] = { - var index = 0 - val combinedAggregate = new Array[Double](binAggregateLength) - while (index < binAggregateLength) { - combinedAggregate(index) = agg1(index) + agg2(index) - index += 1 - } - combinedAggregate - } - // Calculate bin aggregates. timer.start("aggregation") - val binAggregates = { - input.treeAggregate(Array.fill[Double](binAggregateLength)(0))(binSeqOp, binCombOp) + val binAggregates: DTStatsAggregator = { + val initAgg = new DTStatsAggregator(metadata, numNodes) + input.treeAggregate(initAgg)(binSeqOp, DTStatsAggregator.binCombOp) } timer.stop("aggregation") - logDebug("binAggregates.length = " + binAggregates.length) - /** - * Calculate the information gain for a given (feature, split) based upon left/right aggregates. - * @param leftNodeAgg left node aggregates for this (feature, split) - * @param rightNodeAgg right node aggregate for this (feature, split) - * @param topImpurity impurity of the parent node - * @return information gain and statistics for all splits - */ - def calculateGainForSplit( - leftNodeAgg: Array[Double], - rightNodeAgg: Array[Double], - topImpurity: Double): InformationGainStats = { - if (metadata.isClassification) { - val leftTotalCount = leftNodeAgg.sum - val rightTotalCount = rightNodeAgg.sum - - val impurity = { - if (level > 0) { - topImpurity - } else { - // Calculate impurity for root node. - val rootNodeCounts = new Array[Double](numClasses) - var classIndex = 0 - while (classIndex < numClasses) { - rootNodeCounts(classIndex) = leftNodeAgg(classIndex) + rightNodeAgg(classIndex) - classIndex += 1 - } - metadata.impurity.calculate(rootNodeCounts, leftTotalCount + rightTotalCount) - } - } - - val totalCount = leftTotalCount + rightTotalCount - if (totalCount == 0) { - // Return arbitrary prediction. - return new InformationGainStats(0, topImpurity, topImpurity, topImpurity, 0) - } - - // Sum of count for each label - val leftrightNodeAgg: Array[Double] = - leftNodeAgg.zip(rightNodeAgg).map { case (leftCount, rightCount) => - leftCount + rightCount - } - - def indexOfLargestArrayElement(array: Array[Double]): Int = { - val result = array.foldLeft(-1, Double.MinValue, 0) { - case ((maxIndex, maxValue, currentIndex), currentValue) => - if (currentValue > maxValue) { - (currentIndex, currentValue, currentIndex + 1) - } else { - (maxIndex, maxValue, currentIndex + 1) - } - } - if (result._1 < 0) { - throw new RuntimeException("DecisionTree internal error:" + - " calculateGainForSplit failed in indexOfLargestArrayElement") - } - result._1 - } - - val predict = indexOfLargestArrayElement(leftrightNodeAgg) - val prob = leftrightNodeAgg(predict) / totalCount - - val leftImpurity = if (leftTotalCount == 0) { - topImpurity - } else { - metadata.impurity.calculate(leftNodeAgg, leftTotalCount) - } - val rightImpurity = if (rightTotalCount == 0) { - topImpurity - } else { - metadata.impurity.calculate(rightNodeAgg, rightTotalCount) - } - - val leftWeight = leftTotalCount / totalCount - val rightWeight = rightTotalCount / totalCount - - val gain = impurity - leftWeight * leftImpurity - rightWeight * rightImpurity - - new InformationGainStats(gain, impurity, leftImpurity, rightImpurity, predict, prob) - - } else { - // Regression - - val leftCount = leftNodeAgg(0) - val leftSum = leftNodeAgg(1) - val leftSumSquares = leftNodeAgg(2) + // Calculate best splits for all nodes at a given level + timer.start("chooseSplits") + val bestSplits = new Array[(Split, InformationGainStats)](numNodes) + // Iterating over all nodes at this level + var nodeIndex = 0 + while (nodeIndex < numNodes) { + val nodeImpurity = parentImpurities(globalNodeIndexOffset + nodeIndex) + logDebug("node impurity = " + nodeImpurity) + bestSplits(nodeIndex) = + binsToBestSplit(binAggregates, nodeIndex, nodeImpurity, level, metadata, splits) + logDebug("best split = " + bestSplits(nodeIndex)._1) + nodeIndex += 1 + } + timer.stop("chooseSplits") - val rightCount = rightNodeAgg(0) - val rightSum = rightNodeAgg(1) - val rightSumSquares = rightNodeAgg(2) + bestSplits + } - val impurity = { - if (level > 0) { - topImpurity - } else { - // Calculate impurity for root node. - val count = leftCount + rightCount - val sum = leftSum + rightSum - val sumSquares = leftSumSquares + rightSumSquares - metadata.impurity.calculate(count, sum, sumSquares) - } - } + /** + * Calculate the information gain for a given (feature, split) based upon left/right aggregates. + * @param leftImpurityCalculator left node aggregates for this (feature, split) + * @param rightImpurityCalculator right node aggregate for this (feature, split) + * @param topImpurity impurity of the parent node + * @return information gain and statistics for all splits + */ + private def calculateGainForSplit( + leftImpurityCalculator: ImpurityCalculator, + rightImpurityCalculator: ImpurityCalculator, + topImpurity: Double, + level: Int, + metadata: DecisionTreeMetadata): InformationGainStats = { - if (leftCount == 0) { - return new InformationGainStats(0, topImpurity, Double.MinValue, topImpurity, - rightSum / rightCount) - } - if (rightCount == 0) { - return new InformationGainStats(0, topImpurity, topImpurity, - Double.MinValue, leftSum / leftCount) - } + val leftCount = leftImpurityCalculator.count + val rightCount = rightImpurityCalculator.count - val leftImpurity = metadata.impurity.calculate(leftCount, leftSum, leftSumSquares) - val rightImpurity = metadata.impurity.calculate(rightCount, rightSum, rightSumSquares) + val totalCount = leftCount + rightCount + if (totalCount == 0) { + // Return arbitrary prediction. + return new InformationGainStats(0, topImpurity, topImpurity, topImpurity, 0) + } - val leftWeight = leftCount.toDouble / (leftCount + rightCount) - val rightWeight = rightCount.toDouble / (leftCount + rightCount) + val parentNodeAgg = leftImpurityCalculator.copy + parentNodeAgg.add(rightImpurityCalculator) + // impurity of parent node + val impurity = if (level > 0) { + topImpurity + } else { + parentNodeAgg.calculate() + } - val gain = impurity - leftWeight * leftImpurity - rightWeight * rightImpurity + val predict = parentNodeAgg.predict + val prob = parentNodeAgg.prob(predict) - val predict = (leftSum + rightSum) / (leftCount + rightCount) - new InformationGainStats(gain, impurity, leftImpurity, rightImpurity, predict) - } - } + val leftImpurity = leftImpurityCalculator.calculate() // Note: This equals 0 if count = 0 + val rightImpurity = rightImpurityCalculator.calculate() - /** - * Extracts left and right split aggregates. - * @param binData Aggregate array slice from getBinDataForNode. - * For classification: - * For unordered features, this is leftChildData ++ rightChildData, - * each of which is indexed by (feature, split/bin, class), - * with class being the least significant bit. - * For ordered features, this is of size numClasses * numBins * numFeatures. - * For regression: - * This is of size 2 * numFeatures * numBins. - * @return (leftNodeAgg, rightNodeAgg) pair of arrays. - * For classification, each array is of size (numFeatures, (numBins - 1), numClasses). - * For regression, each array is of size (numFeatures, (numBins - 1), 3). - * - */ - def extractLeftRightNodeAggregates( - binData: Array[Double]): (Array[Array[Array[Double]]], Array[Array[Array[Double]]]) = { - - - /** - * The input binData is indexed as (feature, bin, class). - * This computes cumulative sums over splits. - * Each (feature, class) pair is handled separately. - * Note: numSplits = numBins - 1. - * @param leftNodeAgg Each (feature, class) slice is an array over splits. - * Element i (i = 0, ..., numSplits - 2) is set to be - * the cumulative sum (from left) over binData for bins 0, ..., i. - * @param rightNodeAgg Each (feature, class) slice is an array over splits. - * Element i (i = 1, ..., numSplits - 1) is set to be - * the cumulative sum (from right) over binData for bins - * numBins - 1, ..., numBins - 1 - i. - */ - def findAggForOrderedFeatureClassification( - leftNodeAgg: Array[Array[Array[Double]]], - rightNodeAgg: Array[Array[Array[Double]]], - featureIndex: Int) { - - // shift for this featureIndex - val shift = numClasses * featureIndex * numBins - - var classIndex = 0 - while (classIndex < numClasses) { - // left node aggregate for the lowest split - leftNodeAgg(featureIndex)(0)(classIndex) = binData(shift + classIndex) - // right node aggregate for the highest split - rightNodeAgg(featureIndex)(numBins - 2)(classIndex) - = binData(shift + (numClasses * (numBins - 1)) + classIndex) - classIndex += 1 - } + val leftWeight = leftCount / totalCount.toDouble + val rightWeight = rightCount / totalCount.toDouble - // Iterate over all splits. - var splitIndex = 1 - while (splitIndex < numBins - 1) { - // calculating left node aggregate for a split as a sum of left node aggregate of a - // lower split and the left bin aggregate of a bin where the split is a high split - var innerClassIndex = 0 - while (innerClassIndex < numClasses) { - leftNodeAgg(featureIndex)(splitIndex)(innerClassIndex) - = binData(shift + numClasses * splitIndex + innerClassIndex) + - leftNodeAgg(featureIndex)(splitIndex - 1)(innerClassIndex) - rightNodeAgg(featureIndex)(numBins - 2 - splitIndex)(innerClassIndex) = - binData(shift + (numClasses * (numBins - 1 - splitIndex) + innerClassIndex)) + - rightNodeAgg(featureIndex)(numBins - 1 - splitIndex)(innerClassIndex) - innerClassIndex += 1 - } - splitIndex += 1 - } - } + val gain = impurity - leftWeight * leftImpurity - rightWeight * rightImpurity - /** - * Reshape binData for this feature. - * Indexes binData as (feature, split, class) with class as the least significant bit. - * @param leftNodeAgg leftNodeAgg(featureIndex)(splitIndex)(classIndex) = aggregate value - */ - def findAggForUnorderedFeatureClassification( - leftNodeAgg: Array[Array[Array[Double]]], - rightNodeAgg: Array[Array[Array[Double]]], - featureIndex: Int) { - - val rightChildShift = numClasses * numBins * numFeatures - var splitIndex = 0 - while (splitIndex < numBins - 1) { - var classIndex = 0 - while (classIndex < numClasses) { - // shift for this featureIndex - val shift = numClasses * featureIndex * numBins + splitIndex * numClasses - val leftBinValue = binData(shift + classIndex) - val rightBinValue = binData(rightChildShift + shift + classIndex) - leftNodeAgg(featureIndex)(splitIndex)(classIndex) = leftBinValue - rightNodeAgg(featureIndex)(splitIndex)(classIndex) = rightBinValue - classIndex += 1 - } - splitIndex += 1 - } - } + new InformationGainStats(gain, impurity, leftImpurity, rightImpurity, predict, prob) + } - def findAggForRegression( - leftNodeAgg: Array[Array[Array[Double]]], - rightNodeAgg: Array[Array[Array[Double]]], - featureIndex: Int) { - - // shift for this featureIndex - val shift = 3 * featureIndex * numBins - // left node aggregate for the lowest split - leftNodeAgg(featureIndex)(0)(0) = binData(shift + 0) - leftNodeAgg(featureIndex)(0)(1) = binData(shift + 1) - leftNodeAgg(featureIndex)(0)(2) = binData(shift + 2) - - // right node aggregate for the highest split - rightNodeAgg(featureIndex)(numBins - 2)(0) = - binData(shift + (3 * (numBins - 1))) - rightNodeAgg(featureIndex)(numBins - 2)(1) = - binData(shift + (3 * (numBins - 1)) + 1) - rightNodeAgg(featureIndex)(numBins - 2)(2) = - binData(shift + (3 * (numBins - 1)) + 2) - - // Iterate over all splits. - var splitIndex = 1 - while (splitIndex < numBins - 1) { - var i = 0 // index for regression histograms - while (i < 3) { // count, sum, sum^2 - // calculating left node aggregate for a split as a sum of left node aggregate of a - // lower split and the left bin aggregate of a bin where the split is a high split - leftNodeAgg(featureIndex)(splitIndex)(i) = binData(shift + 3 * splitIndex + i) + - leftNodeAgg(featureIndex)(splitIndex - 1)(i) - // calculating right node aggregate for a split as a sum of right node aggregate of a - // higher split and the right bin aggregate of a bin where the split is a low split - rightNodeAgg(featureIndex)(numBins - 2 - splitIndex)(i) = - binData(shift + (3 * (numBins - 1 - splitIndex) + i)) + - rightNodeAgg(featureIndex)(numBins - 1 - splitIndex)(i) - i += 1 - } - splitIndex += 1 - } - } + /** + * Find the best split for a node. + * @param binAggregates Bin statistics. + * @param nodeIndex Index for node to split in this (level, group). + * @param nodeImpurity Impurity of the node (nodeIndex). + * @return tuple for best split: (Split, information gain) + */ + private def binsToBestSplit( + binAggregates: DTStatsAggregator, + nodeIndex: Int, + nodeImpurity: Double, + level: Int, + metadata: DecisionTreeMetadata, + splits: Array[Array[Split]]): (Split, InformationGainStats) = { - if (metadata.isClassification) { - // Initialize left and right split aggregates. - val leftNodeAgg = Array.ofDim[Double](numFeatures, numBins - 1, numClasses) - val rightNodeAgg = Array.ofDim[Double](numFeatures, numBins - 1, numClasses) - var featureIndex = 0 - while (featureIndex < numFeatures) { - if (metadata.isUnordered(featureIndex)) { - findAggForUnorderedFeatureClassification(leftNodeAgg, rightNodeAgg, featureIndex) - } else { - findAggForOrderedFeatureClassification(leftNodeAgg, rightNodeAgg, featureIndex) - } - featureIndex += 1 - } - (leftNodeAgg, rightNodeAgg) - } else { - // Regression - // Initialize left and right split aggregates. - val leftNodeAgg = Array.ofDim[Double](numFeatures, numBins - 1, 3) - val rightNodeAgg = Array.ofDim[Double](numFeatures, numBins - 1, 3) - // Iterate over all features. - var featureIndex = 0 - while (featureIndex < numFeatures) { - findAggForRegression(leftNodeAgg, rightNodeAgg, featureIndex) - featureIndex += 1 - } - (leftNodeAgg, rightNodeAgg) - } - } + logDebug("node impurity = " + nodeImpurity) - /** - * Calculates information gain for all nodes splits. - */ - def calculateGainsForAllNodeSplits( - leftNodeAgg: Array[Array[Array[Double]]], - rightNodeAgg: Array[Array[Array[Double]]], - nodeImpurity: Double): Array[Array[InformationGainStats]] = { - val gains = Array.ofDim[InformationGainStats](numFeatures, numBins - 1) - - var featureIndex = 0 - while (featureIndex < numFeatures) { - val numSplitsForFeature = getNumSplitsForFeature(featureIndex) + // For each (feature, split), calculate the gain, and select the best (feature, split). + Range(0, metadata.numFeatures).map { featureIndex => + val numSplits = metadata.numSplits(featureIndex) + if (metadata.isContinuous(featureIndex)) { + // Cumulative sum (scanLeft) of bin statistics. + // Afterwards, binAggregates for a bin is the sum of aggregates for + // that bin + all preceding bins. + val nodeFeatureOffset = binAggregates.getNodeFeatureOffset(nodeIndex, featureIndex) var splitIndex = 0 - while (splitIndex < numSplitsForFeature) { - gains(featureIndex)(splitIndex) = - calculateGainForSplit(leftNodeAgg(featureIndex)(splitIndex), - rightNodeAgg(featureIndex)(splitIndex), nodeImpurity) + while (splitIndex < numSplits) { + binAggregates.mergeForNodeFeature(nodeFeatureOffset, splitIndex + 1, splitIndex) splitIndex += 1 } - featureIndex += 1 - } - gains - } - - /** - * Get the number of splits for a feature. - */ - def getNumSplitsForFeature(featureIndex: Int): Int = { - if (metadata.isContinuous(featureIndex)) { - numBins - 1 + // Find best split. + val (bestFeatureSplitIndex, bestFeatureGainStats) = + Range(0, numSplits).map { case splitIdx => + val leftChildStats = binAggregates.getImpurityCalculator(nodeFeatureOffset, splitIdx) + val rightChildStats = binAggregates.getImpurityCalculator(nodeFeatureOffset, numSplits) + rightChildStats.subtract(leftChildStats) + val gainStats = + calculateGainForSplit(leftChildStats, rightChildStats, nodeImpurity, level, metadata) + (splitIdx, gainStats) + }.maxBy(_._2.gain) + (splits(featureIndex)(bestFeatureSplitIndex), bestFeatureGainStats) + } else if (metadata.isUnordered(featureIndex)) { + // Unordered categorical feature + val (leftChildOffset, rightChildOffset) = + binAggregates.getLeftRightNodeFeatureOffsets(nodeIndex, featureIndex) + val (bestFeatureSplitIndex, bestFeatureGainStats) = + Range(0, numSplits).map { splitIndex => + val leftChildStats = binAggregates.getImpurityCalculator(leftChildOffset, splitIndex) + val rightChildStats = binAggregates.getImpurityCalculator(rightChildOffset, splitIndex) + val gainStats = + calculateGainForSplit(leftChildStats, rightChildStats, nodeImpurity, level, metadata) + (splitIndex, gainStats) + }.maxBy(_._2.gain) + (splits(featureIndex)(bestFeatureSplitIndex), bestFeatureGainStats) } else { - // Categorical feature - val featureCategories = metadata.featureArity(featureIndex) - if (metadata.isUnordered(featureIndex)) { - (1 << featureCategories - 1) - 1 - } else { - featureCategories - } - } - } - - /** - * Find the best split for a node. - * @param binData Bin data slice for this node, given by getBinDataForNode. - * @param nodeImpurity impurity of the top node - * @return tuple of split and information gain - */ - def binsToBestSplit( - binData: Array[Double], - nodeImpurity: Double): (Split, InformationGainStats) = { - - logDebug("node impurity = " + nodeImpurity) - - // Extract left right node aggregates. - val (leftNodeAgg, rightNodeAgg) = extractLeftRightNodeAggregates(binData) - - // Calculate gains for all splits. - val gains = calculateGainsForAllNodeSplits(leftNodeAgg, rightNodeAgg, nodeImpurity) - - val (bestFeatureIndex, bestSplitIndex, gainStats) = { - // Initialize with infeasible values. - var bestFeatureIndex = Int.MinValue - var bestSplitIndex = Int.MinValue - var bestGainStats = new InformationGainStats(Double.MinValue, -1.0, -1.0, -1.0, -1.0) - // Iterate over features. - var featureIndex = 0 - while (featureIndex < numFeatures) { - // Iterate over all splits. - var splitIndex = 0 - val numSplitsForFeature = getNumSplitsForFeature(featureIndex) - while (splitIndex < numSplitsForFeature) { - val gainStats = gains(featureIndex)(splitIndex) - if (gainStats.gain > bestGainStats.gain) { - bestGainStats = gainStats - bestFeatureIndex = featureIndex - bestSplitIndex = splitIndex + // Ordered categorical feature + val nodeFeatureOffset = binAggregates.getNodeFeatureOffset(nodeIndex, featureIndex) + val numBins = metadata.numBins(featureIndex) + + /* 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 centroidForCategories = if (metadata.isMulticlass) { + // For categorical variables in multiclass classification, + // the bins are ordered by the impurity of their corresponding labels. + Range(0, numBins).map { case featureValue => + val categoryStats = binAggregates.getImpurityCalculator(nodeFeatureOffset, featureValue) + val centroid = if (categoryStats.count != 0) { + categoryStats.calculate() + } else { + Double.MaxValue } - splitIndex += 1 + (featureValue, centroid) + } + } else { // regression or binary classification + // For categorical variables in regression and binary classification, + // the bins are ordered by the centroid of their corresponding labels. + Range(0, numBins).map { case featureValue => + val categoryStats = binAggregates.getImpurityCalculator(nodeFeatureOffset, featureValue) + val centroid = if (categoryStats.count != 0) { + categoryStats.predict + } else { + Double.MaxValue + } + (featureValue, centroid) } - featureIndex += 1 } - (bestFeatureIndex, bestSplitIndex, bestGainStats) - } - logDebug("best split = " + splits(bestFeatureIndex)(bestSplitIndex)) - logDebug("best split bin = " + bins(bestFeatureIndex)(bestSplitIndex)) + logDebug("Centroids for categorical variable: " + centroidForCategories.mkString(",")) - (splits(bestFeatureIndex)(bestSplitIndex), gainStats) - } + // bins sorted by centroids + val categoriesSortedByCentroid = centroidForCategories.toList.sortBy(_._2) - /** - * Get bin data for one node. - */ - def getBinDataForNode(node: Int): Array[Double] = { - if (metadata.isClassification) { - if (isMulticlassWithCategoricalFeatures) { - val shift = numClasses * node * numBins * numFeatures - val rightChildShift = numClasses * numBins * numFeatures * numNodes - val binsForNode = { - val leftChildData - = binAggregates.slice(shift, shift + numClasses * numBins * numFeatures) - val rightChildData - = binAggregates.slice(rightChildShift + shift, - rightChildShift + shift + numClasses * numBins * numFeatures) - leftChildData ++ rightChildData - } - binsForNode - } else { - val shift = numClasses * node * numBins * numFeatures - val binsForNode = binAggregates.slice(shift, shift + numClasses * numBins * numFeatures) - binsForNode + logDebug("Sorted centroids for categorical variable = " + + categoriesSortedByCentroid.mkString(",")) + + // Cumulative sum (scanLeft) of bin statistics. + // Afterwards, binAggregates for a bin is the sum of aggregates for + // that bin + all preceding bins. + var splitIndex = 0 + while (splitIndex < numSplits) { + val currentCategory = categoriesSortedByCentroid(splitIndex)._1 + val nextCategory = categoriesSortedByCentroid(splitIndex + 1)._1 + binAggregates.mergeForNodeFeature(nodeFeatureOffset, nextCategory, currentCategory) + splitIndex += 1 } - } else { - // Regression - val shift = 3 * node * numBins * numFeatures - val binsForNode = binAggregates.slice(shift, shift + 3 * numBins * numFeatures) - binsForNode + // lastCategory = index of bin with total aggregates for this (node, feature) + val lastCategory = categoriesSortedByCentroid.last._1 + // Find best split. + val (bestFeatureSplitIndex, bestFeatureGainStats) = + Range(0, numSplits).map { splitIndex => + val featureValue = categoriesSortedByCentroid(splitIndex)._1 + val leftChildStats = + binAggregates.getImpurityCalculator(nodeFeatureOffset, featureValue) + val rightChildStats = + binAggregates.getImpurityCalculator(nodeFeatureOffset, lastCategory) + rightChildStats.subtract(leftChildStats) + val gainStats = + calculateGainForSplit(leftChildStats, rightChildStats, nodeImpurity, level, metadata) + (splitIndex, gainStats) + }.maxBy(_._2.gain) + val categoriesForSplit = + categoriesSortedByCentroid.map(_._1.toDouble).slice(0, bestFeatureSplitIndex + 1) + val bestFeatureSplit = + new Split(featureIndex, Double.MinValue, Categorical, categoriesForSplit) + (bestFeatureSplit, bestFeatureGainStats) } - } - - // Calculate best splits for all nodes at a given level - timer.start("chooseSplits") - val bestSplits = new Array[(Split, InformationGainStats)](numNodes) - // Iterating over all nodes at this level - var node = 0 - while (node < numNodes) { - val nodeImpurityIndex = (1 << level) - 1 + node + groupShift - val binsForNode: Array[Double] = getBinDataForNode(node) - logDebug("nodeImpurityIndex = " + nodeImpurityIndex) - val parentNodeImpurity = parentImpurities(nodeImpurityIndex) - logDebug("parent node impurity = " + parentNodeImpurity) - bestSplits(node) = binsToBestSplit(binsForNode, parentNodeImpurity) - node += 1 - } - timer.stop("chooseSplits") - - bestSplits + }.maxBy(_._2.gain) } /** * Get the number of values to be stored per node in the bin aggregates. - * - * @param numBins Number of bins = 1 + number of possible splits. */ - private def getElementsPerNode(metadata: DecisionTreeMetadata, numBins: Int): Int = { + private def getElementsPerNode(metadata: DecisionTreeMetadata): Int = { + val totalBins = metadata.numBins.sum if (metadata.isClassification) { - if (metadata.isMulticlassWithCategoricalFeatures) { - 2 * metadata.numClasses * numBins * metadata.numFeatures - } else { - metadata.numClasses * numBins * metadata.numFeatures - } + metadata.numClasses * totalBins } else { - 3 * numBins * metadata.numFeatures + 3 * totalBins } } @@ -1284,6 +919,7 @@ object DecisionTree extends Serializable with Logging { * Continuous features: * For each feature, there are numBins - 1 possible splits representing the possible binary * decisions at each node in the tree. + * This finds locations (feature values) for splits using a subsample of the data. * * Categorical features: * For each feature, there is 1 bin per split. @@ -1292,7 +928,6 @@ object DecisionTree extends Serializable with Logging { * For multiclass classification with a low-arity feature * (i.e., if isMulticlass && isSpaceSufficientForAllCategoricalSplits), * the feature is split based on subsets of categories. - * There are (1 << maxFeatureValue - 1) - 1 splits. * (b) "ordered features" * For regression and binary classification, * and for multiclass classification with a high-arity feature, @@ -1302,7 +937,7 @@ object DecisionTree extends Serializable with Logging { * @param metadata Learning and dataset metadata * @return A tuple of (splits, bins). * Splits is an Array of [[org.apache.spark.mllib.tree.model.Split]] - * of size (numFeatures, numBins - 1). + * of size (numFeatures, numSplits). * Bins is an Array of [[org.apache.spark.mllib.tree.model.Bin]] * of size (numFeatures, numBins). */ @@ -1310,84 +945,80 @@ object DecisionTree extends Serializable with Logging { input: RDD[LabeledPoint], metadata: DecisionTreeMetadata): (Array[Array[Split]], Array[Array[Bin]]) = { - val count = input.count() + logDebug("isMulticlass = " + metadata.isMulticlass) - // Find the number of features by looking at the first sample - val numFeatures = input.take(1)(0).features.size - - val maxBins = metadata.maxBins - val numBins = if (maxBins <= count) maxBins else count.toInt - logDebug("numBins = " + numBins) - val isMulticlass = metadata.isMulticlass - logDebug("isMulticlass = " + isMulticlass) - - /* - * Ensure numBins is always greater than the categories. For multiclass classification, - * numBins should be greater than 2^(maxCategories - 1) - 1. - * It's a limitation of the current implementation but a reasonable trade-off since features - * with large number of categories get favored over continuous features. - * - * This needs to be checked here instead of in Strategy since numBins can be determined - * by the number of training examples. - * TODO: Allow this case, where we simply will know nothing about some categories. - */ - if (metadata.featureArity.size > 0) { - val maxCategoriesForFeatures = metadata.featureArity.maxBy(_._2)._2 - require(numBins > maxCategoriesForFeatures, "numBins should be greater than max categories " + - "in categorical features") - } - - // Calculate the number of sample for approximate quantile calculation. - val requiredSamples = numBins*numBins - val fraction = if (requiredSamples < count) requiredSamples.toDouble / count else 1.0 - logDebug("fraction of data used for calculating quantiles = " + fraction) + val numFeatures = metadata.numFeatures - // sampled input for RDD calculation - val sampledInput = + // Sample the input only if there are continuous features. + val hasContinuousFeatures = Range(0, numFeatures).exists(metadata.isContinuous) + val sampledInput = if (hasContinuousFeatures) { + // Calculate the number of samples for approximate quantile calculation. + val requiredSamples = math.max(metadata.maxBins * metadata.maxBins, 10000) + val fraction = if (requiredSamples < metadata.numExamples) { + requiredSamples.toDouble / metadata.numExamples + } else { + 1.0 + } + logDebug("fraction of data used for calculating quantiles = " + fraction) input.sample(withReplacement = false, fraction, new XORShiftRandom().nextInt()).collect() - val numSamples = sampledInput.length - - val stride: Double = numSamples.toDouble / numBins - logDebug("stride = " + stride) + } else { + new Array[LabeledPoint](0) + } metadata.quantileStrategy match { case Sort => - val splits = Array.ofDim[Split](numFeatures, numBins - 1) - val bins = Array.ofDim[Bin](numFeatures, numBins) + val splits = new Array[Array[Split]](numFeatures) + val bins = new Array[Array[Bin]](numFeatures) // Find all splits. - // Iterate over all features. var featureIndex = 0 while (featureIndex < numFeatures) { - // Check whether the feature is continuous. - val isFeatureContinuous = metadata.isContinuous(featureIndex) - if (isFeatureContinuous) { + val numSplits = metadata.numSplits(featureIndex) + val numBins = metadata.numBins(featureIndex) + if (metadata.isContinuous(featureIndex)) { + val numSamples = sampledInput.length + splits(featureIndex) = new Array[Split](numSplits) + bins(featureIndex) = new Array[Bin](numBins) val featureSamples = sampledInput.map(lp => lp.features(featureIndex)).sorted - val stride: Double = numSamples.toDouble / numBins + val stride: Double = numSamples.toDouble / metadata.numBins(featureIndex) logDebug("stride = " + stride) - for (index <- 0 until numBins - 1) { - val sampleIndex = index * stride.toInt + for (splitIndex <- 0 until numSplits) { + val sampleIndex = splitIndex * stride.toInt // Set threshold halfway in between 2 samples. val threshold = (featureSamples(sampleIndex) + featureSamples(sampleIndex + 1)) / 2.0 - val split = new Split(featureIndex, threshold, Continuous, List()) - splits(featureIndex)(index) = split + splits(featureIndex)(splitIndex) = + new Split(featureIndex, threshold, Continuous, List()) } - } else { // Categorical feature - val featureCategories = metadata.featureArity(featureIndex) - - // Use different bin/split calculation strategy for categorical features in multiclass - // classification that satisfy the space constraint. + bins(featureIndex)(0) = new Bin(new DummyLowSplit(featureIndex, Continuous), + splits(featureIndex)(0), Continuous, Double.MinValue) + for (splitIndex <- 1 until numSplits) { + bins(featureIndex)(splitIndex) = + new Bin(splits(featureIndex)(splitIndex - 1), splits(featureIndex)(splitIndex), + Continuous, Double.MinValue) + } + bins(featureIndex)(numSplits) = new Bin(splits(featureIndex)(numSplits - 1), + new DummyHighSplit(featureIndex, Continuous), Continuous, Double.MinValue) + } else { + // Categorical feature + val featureArity = metadata.featureArity(featureIndex) if (metadata.isUnordered(featureIndex)) { - // 2^(maxFeatureValue- 1) - 1 combinations - var index = 0 - while (index < (1 << featureCategories - 1) - 1) { - val categories: List[Double] - = extractMultiClassCategories(index + 1, featureCategories) - splits(featureIndex)(index) - = new Split(featureIndex, Double.MinValue, Categorical, categories) - bins(featureIndex)(index) = { - if (index == 0) { + // TODO: The second half of the bins are unused. Actually, we could just use + // splits and not build bins for unordered features. That should be part of + // a later PR since it will require changing other code (using splits instead + // of bins in a few places). + // Unordered features + // 2^(maxFeatureValue - 1) - 1 combinations + splits(featureIndex) = new Array[Split](numSplits) + bins(featureIndex) = new Array[Bin](numBins) + var splitIndex = 0 + while (splitIndex < numSplits) { + val categories: List[Double] = + extractMultiClassCategories(splitIndex + 1, featureArity) + splits(featureIndex)(splitIndex) = + new Split(featureIndex, Double.MinValue, Categorical, categories) + bins(featureIndex)(splitIndex) = { + if (splitIndex == 0) { new Bin( new DummyCategoricalSplit(featureIndex, Categorical), splits(featureIndex)(0), @@ -1395,96 +1026,24 @@ object DecisionTree extends Serializable with Logging { Double.MinValue) } else { new Bin( - splits(featureIndex)(index - 1), - splits(featureIndex)(index), + splits(featureIndex)(splitIndex - 1), + splits(featureIndex)(splitIndex), Categorical, Double.MinValue) } } - index += 1 - } - } else { // ordered feature - /* For a given categorical feature, use a subsample of the data - * to choose how to arrange possible splits. - * This examines each category and computes a centroid. - * These centroids are later used to sort the possible splits. - * centroidForCategories is a mapping: category (for the given feature) --> centroid - */ - val centroidForCategories = { - if (isMulticlass) { - // For categorical variables in multiclass classification, - // each bin is a category. The bins are sorted and they - // are ordered by calculating the impurity of their corresponding labels. - sampledInput.map(lp => (lp.features(featureIndex), lp.label)) - .groupBy(_._1) - .mapValues(x => x.groupBy(_._2).mapValues(x => x.size.toDouble)) - .map(x => (x._1, x._2.values.toArray)) - .map(x => (x._1, metadata.impurity.calculate(x._2, x._2.sum))) - } else { // regression or binary classification - // For categorical variables in regression and binary classification, - // each bin is a category. The bins are sorted and they - // are ordered by calculating the centroid of their corresponding labels. - sampledInput.map(lp => (lp.features(featureIndex), lp.label)) - .groupBy(_._1) - .mapValues(x => x.map(_._2).sum / x.map(_._1).length) - } - } - - logDebug("centroid for categories = " + centroidForCategories.mkString(",")) - - // Check for missing categorical variables and putting them last in the sorted list. - val fullCentroidForCategories = scala.collection.mutable.Map[Double,Double]() - for (i <- 0 until featureCategories) { - if (centroidForCategories.contains(i)) { - fullCentroidForCategories(i) = centroidForCategories(i) - } else { - fullCentroidForCategories(i) = Double.MaxValue - } - } - - // bins sorted by centroids - val categoriesSortedByCentroid = fullCentroidForCategories.toList.sortBy(_._2) - - logDebug("centroid for categorical variable = " + categoriesSortedByCentroid) - - var categoriesForSplit = List[Double]() - categoriesSortedByCentroid.iterator.zipWithIndex.foreach { - case ((key, value), index) => - categoriesForSplit = key :: categoriesForSplit - splits(featureIndex)(index) = new Split(featureIndex, Double.MinValue, - Categorical, categoriesForSplit) - bins(featureIndex)(index) = { - if (index == 0) { - new Bin(new DummyCategoricalSplit(featureIndex, Categorical), - splits(featureIndex)(0), Categorical, key) - } else { - new Bin(splits(featureIndex)(index-1), splits(featureIndex)(index), - Categorical, key) - } - } + splitIndex += 1 } + } else { + // Ordered features + // Bins correspond to feature values, so we do not need to compute splits or bins + // beforehand. Splits are constructed as needed during training. + splits(featureIndex) = new Array[Split](0) + bins(featureIndex) = new Array[Bin](0) } } featureIndex += 1 } - - // Find all bins. - featureIndex = 0 - while (featureIndex < numFeatures) { - val isFeatureContinuous = metadata.isContinuous(featureIndex) - if (isFeatureContinuous) { // Bins for categorical variables are already assigned. - bins(featureIndex)(0) = new Bin(new DummyLowSplit(featureIndex, Continuous), - splits(featureIndex)(0), Continuous, Double.MinValue) - for (index <- 1 until numBins - 1) { - val bin = new Bin(splits(featureIndex)(index-1), splits(featureIndex)(index), - Continuous, Double.MinValue) - bins(featureIndex)(index) = bin - } - bins(featureIndex)(numBins-1) = new Bin(splits(featureIndex)(numBins-2), - new DummyHighSplit(featureIndex, Continuous), Continuous, Double.MinValue) - } - featureIndex += 1 - } (splits, bins) case MinMax => throw new UnsupportedOperationException("minmax not supported yet.") diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/DTStatsAggregator.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/DTStatsAggregator.scala new file mode 100644 index 0000000000..866d85a79b --- /dev/null +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/DTStatsAggregator.scala @@ -0,0 +1,213 @@ +/* + * 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.impl + +import org.apache.spark.mllib.tree.impurity._ + +/** + * DecisionTree statistics aggregator. + * This holds a flat array of statistics for a set of (nodes, features, bins) + * and helps with indexing. + */ +private[tree] class DTStatsAggregator( + val metadata: DecisionTreeMetadata, + val numNodes: Int) extends Serializable { + + /** + * [[ImpurityAggregator]] instance specifying the impurity type. + */ + val impurityAggregator: ImpurityAggregator = metadata.impurity match { + case Gini => new GiniAggregator(metadata.numClasses) + case Entropy => new EntropyAggregator(metadata.numClasses) + case Variance => new VarianceAggregator() + case _ => throw new IllegalArgumentException(s"Bad impurity parameter: ${metadata.impurity}") + } + + /** + * Number of elements (Double values) used for the sufficient statistics of each bin. + */ + val statsSize: Int = impurityAggregator.statsSize + + val numFeatures: Int = metadata.numFeatures + + /** + * Number of bins for each feature. This is indexed by the feature index. + */ + val numBins: Array[Int] = metadata.numBins + + /** + * Number of splits for the given feature. + */ + def numSplits(featureIndex: Int): Int = metadata.numSplits(featureIndex) + + /** + * Indicator for each feature of whether that feature is an unordered feature. + * TODO: Is Array[Boolean] any faster? + */ + def isUnordered(featureIndex: Int): Boolean = metadata.isUnordered(featureIndex) + + /** + * Offset for each feature for calculating indices into the [[allStats]] array. + */ + private val featureOffsets: Array[Int] = { + def featureOffsetsCalc(total: Int, featureIndex: Int): Int = { + if (isUnordered(featureIndex)) { + total + 2 * numBins(featureIndex) + } else { + total + numBins(featureIndex) + } + } + Range(0, numFeatures).scanLeft(0)(featureOffsetsCalc).map(statsSize * _).toArray + } + + /** + * Number of elements for each node, corresponding to stride between nodes in [[allStats]]. + */ + private val nodeStride: Int = featureOffsets.last + + /** + * Total number of elements stored in this aggregator. + */ + val allStatsSize: Int = numNodes * nodeStride + + /** + * Flat array of elements. + * Index for start of stats for a (node, feature, bin) is: + * index = nodeIndex * nodeStride + featureOffsets(featureIndex) + binIndex * statsSize + * Note: For unordered features, the left child stats have binIndex in [0, numBins(featureIndex)) + * and the right child stats in [numBins(featureIndex), 2 * numBins(featureIndex)) + */ + val allStats: Array[Double] = new Array[Double](allStatsSize) + + /** + * Get an [[ImpurityCalculator]] for a given (node, feature, bin). + * @param nodeFeatureOffset For ordered features, this is a pre-computed (node, feature) offset + * from [[getNodeFeatureOffset]]. + * For unordered features, this is a pre-computed + * (node, feature, left/right child) offset from + * [[getLeftRightNodeFeatureOffsets]]. + */ + def getImpurityCalculator(nodeFeatureOffset: Int, binIndex: Int): ImpurityCalculator = { + impurityAggregator.getCalculator(allStats, nodeFeatureOffset + binIndex * statsSize) + } + + /** + * Update the stats for a given (node, feature, bin) for ordered features, using the given label. + */ + def update(nodeIndex: Int, featureIndex: Int, binIndex: Int, label: Double): Unit = { + val i = nodeIndex * nodeStride + featureOffsets(featureIndex) + binIndex * statsSize + impurityAggregator.update(allStats, i, label) + } + + /** + * Pre-compute node offset for use with [[nodeUpdate]]. + */ + def getNodeOffset(nodeIndex: Int): Int = nodeIndex * nodeStride + + /** + * Faster version of [[update]]. + * Update the stats for a given (node, feature, bin) for ordered features, using the given label. + * @param nodeOffset Pre-computed node offset from [[getNodeOffset]]. + */ + def nodeUpdate(nodeOffset: Int, featureIndex: Int, binIndex: Int, label: Double): Unit = { + val i = nodeOffset + featureOffsets(featureIndex) + binIndex * statsSize + impurityAggregator.update(allStats, i, label) + } + + /** + * Pre-compute (node, feature) offset for use with [[nodeFeatureUpdate]]. + * For ordered features only. + */ + def getNodeFeatureOffset(nodeIndex: Int, featureIndex: Int): Int = { + require(!isUnordered(featureIndex), + s"DTStatsAggregator.getNodeFeatureOffset is for ordered features only, but was called" + + s" for unordered feature $featureIndex.") + nodeIndex * nodeStride + featureOffsets(featureIndex) + } + + /** + * Pre-compute (node, feature) offset for use with [[nodeFeatureUpdate]]. + * For unordered features only. + */ + def getLeftRightNodeFeatureOffsets(nodeIndex: Int, featureIndex: Int): (Int, Int) = { + require(isUnordered(featureIndex), + s"DTStatsAggregator.getLeftRightNodeFeatureOffsets is for unordered features only," + + s" but was called for ordered feature $featureIndex.") + val baseOffset = nodeIndex * nodeStride + featureOffsets(featureIndex) + (baseOffset, baseOffset + numBins(featureIndex) * statsSize) + } + + /** + * Faster version of [[update]]. + * Update the stats for a given (node, feature, bin), using the given label. + * @param nodeFeatureOffset For ordered features, this is a pre-computed (node, feature) offset + * from [[getNodeFeatureOffset]]. + * For unordered features, this is a pre-computed + * (node, feature, left/right child) offset from + * [[getLeftRightNodeFeatureOffsets]]. + */ + def nodeFeatureUpdate(nodeFeatureOffset: Int, binIndex: Int, label: Double): Unit = { + impurityAggregator.update(allStats, nodeFeatureOffset + binIndex * statsSize, label) + } + + /** + * For a given (node, feature), merge the stats for two bins. + * @param nodeFeatureOffset For ordered features, this is a pre-computed (node, feature) offset + * from [[getNodeFeatureOffset]]. + * For unordered features, this is a pre-computed + * (node, feature, left/right child) offset from + * [[getLeftRightNodeFeatureOffsets]]. + * @param binIndex The other bin is merged into this bin. + * @param otherBinIndex This bin is not modified. + */ + def mergeForNodeFeature(nodeFeatureOffset: Int, binIndex: Int, otherBinIndex: Int): Unit = { + impurityAggregator.merge(allStats, nodeFeatureOffset + binIndex * statsSize, + nodeFeatureOffset + otherBinIndex * statsSize) + } + + /** + * Merge this aggregator with another, and returns this aggregator. + * This method modifies this aggregator in-place. + */ + def merge(other: DTStatsAggregator): DTStatsAggregator = { + require(allStatsSize == other.allStatsSize, + s"DTStatsAggregator.merge requires that both aggregators have the same length stats vectors." + + s" This aggregator is of length $allStatsSize, but the other is ${other.allStatsSize}.") + var i = 0 + // TODO: Test BLAS.axpy + while (i < allStatsSize) { + allStats(i) += other.allStats(i) + i += 1 + } + this + } + +} + +private[tree] object DTStatsAggregator extends Serializable { + + /** + * Combines two aggregates (modifying the first) and returns the combination. + */ + def binCombOp( + agg1: DTStatsAggregator, + agg2: DTStatsAggregator): DTStatsAggregator = { + agg1.merge(agg2) + } + +} diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/DecisionTreeMetadata.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/DecisionTreeMetadata.scala index d9eda354dc..e95add7558 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/DecisionTreeMetadata.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/DecisionTreeMetadata.scala @@ -26,14 +26,15 @@ import org.apache.spark.mllib.tree.configuration.Strategy import org.apache.spark.mllib.tree.impurity.Impurity import org.apache.spark.rdd.RDD - /** * Learning and dataset metadata for DecisionTree. * * @param numClasses For classification: labels can take values {0, ..., numClasses - 1}. * For regression: fixed at 0 (no meaning). + * @param maxBins Maximum number of bins, for all features. * @param featureArity Map: categorical feature index --> arity. * I.e., the feature takes values in {0, ..., arity - 1}. + * @param numBins Number of bins for each feature. */ private[tree] class DecisionTreeMetadata( val numFeatures: Int, @@ -42,6 +43,7 @@ private[tree] class DecisionTreeMetadata( val maxBins: Int, val featureArity: Map[Int, Int], val unorderedFeatures: Set[Int], + val numBins: Array[Int], val impurity: Impurity, val quantileStrategy: QuantileStrategy) extends Serializable { @@ -57,10 +59,26 @@ private[tree] class DecisionTreeMetadata( def isContinuous(featureIndex: Int): Boolean = !featureArity.contains(featureIndex) + /** + * Number of splits for the given feature. + * For unordered features, there are 2 bins per split. + * For ordered features, there is 1 more bin than split. + */ + def numSplits(featureIndex: Int): Int = if (isUnordered(featureIndex)) { + numBins(featureIndex) >> 1 + } else { + numBins(featureIndex) - 1 + } + } private[tree] object DecisionTreeMetadata { + /** + * Construct a [[DecisionTreeMetadata]] instance for this dataset and parameters. + * This computes which categorical features will be ordered vs. unordered, + * as well as the number of splits and bins for each feature. + */ def buildMetadata(input: RDD[LabeledPoint], strategy: Strategy): DecisionTreeMetadata = { val numFeatures = input.take(1)(0).features.size @@ -70,32 +88,55 @@ private[tree] object DecisionTreeMetadata { case Regression => 0 } - val maxBins = math.min(strategy.maxBins, numExamples).toInt - val log2MaxBinsp1 = math.log(maxBins + 1) / math.log(2.0) + val maxPossibleBins = math.min(strategy.maxBins, numExamples).toInt + + // We check the number of bins here against maxPossibleBins. + // This needs to be checked here instead of in Strategy since maxPossibleBins can be modified + // based on the number of training examples. + if (strategy.categoricalFeaturesInfo.nonEmpty) { + val maxCategoriesPerFeature = strategy.categoricalFeaturesInfo.values.max + require(maxCategoriesPerFeature <= maxPossibleBins, + s"DecisionTree requires maxBins (= $maxPossibleBins) >= max categories " + + s"in categorical features (= $maxCategoriesPerFeature)") + } val unorderedFeatures = new mutable.HashSet[Int]() + val numBins = Array.fill[Int](numFeatures)(maxPossibleBins) if (numClasses > 2) { - strategy.categoricalFeaturesInfo.foreach { case (f, k) => - if (k - 1 < log2MaxBinsp1) { - // Note: The above check is equivalent to checking: - // numUnorderedBins = (1 << k - 1) - 1 < maxBins - unorderedFeatures.add(f) + // Multiclass classification + val maxCategoriesForUnorderedFeature = + ((math.log(maxPossibleBins / 2 + 1) / math.log(2.0)) + 1).floor.toInt + strategy.categoricalFeaturesInfo.foreach { case (featureIndex, numCategories) => + // Decide if some categorical features should be treated as unordered features, + // which require 2 * ((1 << numCategories - 1) - 1) bins. + // We do this check with log values to prevent overflows in case numCategories is large. + // The next check is equivalent to: 2 * ((1 << numCategories - 1) - 1) <= maxBins + if (numCategories <= maxCategoriesForUnorderedFeature) { + unorderedFeatures.add(featureIndex) + numBins(featureIndex) = numUnorderedBins(numCategories) } else { - // TODO: Allow this case, where we simply will know nothing about some categories? - require(k < maxBins, s"maxBins (= $maxBins) should be greater than max categories " + - s"in categorical features (>= $k)") + numBins(featureIndex) = numCategories } } } else { - strategy.categoricalFeaturesInfo.foreach { case (f, k) => - require(k < maxBins, s"maxBins (= $maxBins) should be greater than max categories " + - s"in categorical features (>= $k)") + // Binary classification or regression + strategy.categoricalFeaturesInfo.foreach { case (featureIndex, numCategories) => + numBins(featureIndex) = numCategories } } - new DecisionTreeMetadata(numFeatures, numExamples, numClasses, maxBins, - strategy.categoricalFeaturesInfo, unorderedFeatures.toSet, + new DecisionTreeMetadata(numFeatures, numExamples, numClasses, numBins.max, + strategy.categoricalFeaturesInfo, unorderedFeatures.toSet, numBins, strategy.impurity, strategy.quantileCalculationStrategy) } + /** + * Given the arity of a categorical feature (arity = number of categories), + * return the number of bins for the feature if it is to be treated as an unordered feature. + * There is 1 split for every partitioning of categories into 2 disjoint, non-empty sets; + * there are math.pow(2, arity - 1) - 1 such splits. + * Each split has 2 corresponding bins. + */ + def numUnorderedBins(arity: Int): Int = 2 * ((1 << arity - 1) - 1) + } diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/TreePoint.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/TreePoint.scala index 170e43e222..35e361ae30 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/TreePoint.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/impl/TreePoint.scala @@ -48,54 +48,63 @@ private[tree] object TreePoint { * binning feature values in preparation for DecisionTree training. * @param input Input dataset. * @param bins Bins for features, of size (numFeatures, numBins). - * @param metadata Learning and dataset metadata + * @param metadata Learning and dataset metadata * @return TreePoint dataset representation */ def convertToTreeRDD( input: RDD[LabeledPoint], bins: Array[Array[Bin]], metadata: DecisionTreeMetadata): RDD[TreePoint] = { + // Construct arrays for featureArity and isUnordered for efficiency in the inner loop. + val featureArity: Array[Int] = new Array[Int](metadata.numFeatures) + val isUnordered: Array[Boolean] = new Array[Boolean](metadata.numFeatures) + var featureIndex = 0 + while (featureIndex < metadata.numFeatures) { + featureArity(featureIndex) = metadata.featureArity.getOrElse(featureIndex, 0) + isUnordered(featureIndex) = metadata.isUnordered(featureIndex) + featureIndex += 1 + } input.map { x => - TreePoint.labeledPointToTreePoint(x, bins, metadata) + TreePoint.labeledPointToTreePoint(x, bins, featureArity, isUnordered) } } /** * Convert one LabeledPoint into its TreePoint representation. * @param bins Bins for features, of size (numFeatures, numBins). + * @param featureArity Array indexed by feature, with value 0 for continuous and numCategories + * for categorical features. + * @param isUnordered Array index by feature, with value true for unordered categorical features. */ private def labeledPointToTreePoint( labeledPoint: LabeledPoint, bins: Array[Array[Bin]], - metadata: DecisionTreeMetadata): TreePoint = { - + featureArity: Array[Int], + isUnordered: Array[Boolean]): TreePoint = { val numFeatures = labeledPoint.features.size - val numBins = bins(0).size val arr = new Array[Int](numFeatures) var featureIndex = 0 while (featureIndex < numFeatures) { - arr(featureIndex) = findBin(featureIndex, labeledPoint, metadata.isContinuous(featureIndex), - metadata.isUnordered(featureIndex), bins, metadata.featureArity) + arr(featureIndex) = findBin(featureIndex, labeledPoint, featureArity(featureIndex), + isUnordered(featureIndex), bins) featureIndex += 1 } - new TreePoint(labeledPoint.label, arr) } /** * Find bin for one (labeledPoint, feature). * + * @param featureArity 0 for continuous features; number of categories for categorical features. * @param isUnorderedFeature (only applies if feature is categorical) * @param bins Bins for features, of size (numFeatures, numBins). - * @param categoricalFeaturesInfo Map over categorical features: feature index --> feature arity */ private def findBin( featureIndex: Int, labeledPoint: LabeledPoint, - isFeatureContinuous: Boolean, + featureArity: Int, isUnorderedFeature: Boolean, - bins: Array[Array[Bin]], - categoricalFeaturesInfo: Map[Int, Int]): Int = { + bins: Array[Array[Bin]]): Int = { /** * Binary search helper method for continuous feature. @@ -121,44 +130,7 @@ private[tree] object TreePoint { -1 } - /** - * Sequential search helper method to find bin for categorical feature in multiclass - * classification. The category is returned since each category can belong to multiple - * splits. The actual left/right child allocation per split is performed in the - * sequential phase of the bin aggregate operation. - */ - def sequentialBinSearchForUnorderedCategoricalFeatureInClassification(): Int = { - labeledPoint.features(featureIndex).toInt - } - - /** - * Sequential search helper method to find bin for categorical feature - * (for classification and regression). - */ - def sequentialBinSearchForOrderedCategoricalFeature(): Int = { - val featureCategories = categoricalFeaturesInfo(featureIndex) - val featureValue = labeledPoint.features(featureIndex) - var binIndex = 0 - while (binIndex < featureCategories) { - val bin = bins(featureIndex)(binIndex) - val categories = bin.highSplit.categories - if (categories.contains(featureValue)) { - return binIndex - } - binIndex += 1 - } - if (featureValue < 0 || featureValue >= featureCategories) { - throw new IllegalArgumentException( - s"DecisionTree given invalid data:" + - s" Feature $featureIndex is categorical with values in" + - s" {0,...,${featureCategories - 1}," + - s" but a data point gives it value $featureValue.\n" + - " Bad data point: " + labeledPoint.toString) - } - -1 - } - - if (isFeatureContinuous) { + if (featureArity == 0) { // Perform binary search for finding bin for continuous features. val binIndex = binarySearchForBins() if (binIndex == -1) { @@ -168,18 +140,17 @@ private[tree] object TreePoint { } binIndex } else { - // Perform sequential search to find bin for categorical features. - val binIndex = if (isUnorderedFeature) { - sequentialBinSearchForUnorderedCategoricalFeatureInClassification() - } else { - sequentialBinSearchForOrderedCategoricalFeature() - } - if (binIndex == -1) { - throw new RuntimeException("No bin was found for categorical feature." + - " This error can occur when given invalid data values (such as NaN)." + - s" Feature index: $featureIndex. Feature value: ${labeledPoint.features(featureIndex)}") + // Categorical feature bins are indexed by feature values. + val featureValue = labeledPoint.features(featureIndex) + if (featureValue < 0 || featureValue >= featureArity) { + throw new IllegalArgumentException( + s"DecisionTree given invalid data:" + + s" Feature $featureIndex is categorical with values in" + + s" {0,...,${featureArity - 1}," + + s" but a data point gives it value $featureValue.\n" + + " Bad data point: " + labeledPoint.toString) } - binIndex + featureValue.toInt } } } diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Entropy.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Entropy.scala index 96d2471e1f..1c8afc2d0f 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Entropy.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Entropy.scala @@ -74,3 +74,87 @@ object Entropy extends Impurity { def instance = this } + +/** + * Class for updating views of a vector of sufficient statistics, + * in order to compute impurity from a sample. + * Note: Instances of this class do not hold the data; they operate on views of the data. + * @param numClasses Number of classes for label. + */ +private[tree] class EntropyAggregator(numClasses: Int) + extends ImpurityAggregator(numClasses) with Serializable { + + /** + * Update stats for one (node, feature, bin) with the given label. + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for this (node, feature, bin). + */ + def update(allStats: Array[Double], offset: Int, label: Double): Unit = { + if (label >= statsSize) { + throw new IllegalArgumentException(s"EntropyAggregator given label $label" + + s" but requires label < numClasses (= $statsSize).") + } + allStats(offset + label.toInt) += 1 + } + + /** + * Get an [[ImpurityCalculator]] for a (node, feature, bin). + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for this (node, feature, bin). + */ + def getCalculator(allStats: Array[Double], offset: Int): EntropyCalculator = { + new EntropyCalculator(allStats.view(offset, offset + statsSize).toArray) + } + +} + +/** + * Stores statistics for one (node, feature, bin) for calculating impurity. + * Unlike [[EntropyAggregator]], this class stores its own data and is for a specific + * (node, feature, bin). + * @param stats Array of sufficient statistics for a (node, feature, bin). + */ +private[tree] class EntropyCalculator(stats: Array[Double]) extends ImpurityCalculator(stats) { + + /** + * Make a deep copy of this [[ImpurityCalculator]]. + */ + def copy: EntropyCalculator = new EntropyCalculator(stats.clone()) + + /** + * Calculate the impurity from the stored sufficient statistics. + */ + def calculate(): Double = Entropy.calculate(stats, stats.sum) + + /** + * Number of data points accounted for in the sufficient statistics. + */ + def count: Long = stats.sum.toLong + + /** + * Prediction which should be made based on the sufficient statistics. + */ + def predict: Double = if (count == 0) { + 0 + } else { + indexOfLargestArrayElement(stats) + } + + /** + * Probability of the label given by [[predict]]. + */ + override def prob(label: Double): Double = { + val lbl = label.toInt + require(lbl < stats.length, + s"EntropyCalculator.prob given invalid label: $lbl (should be < ${stats.length}") + val cnt = count + if (cnt == 0) { + 0 + } else { + stats(lbl) / cnt + } + } + + override def toString: String = s"EntropyCalculator(stats = [${stats.mkString(", ")}])" + +} diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Gini.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Gini.scala index d586f44904..5cfdf345d1 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Gini.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Gini.scala @@ -70,3 +70,87 @@ object Gini extends Impurity { def instance = this } + +/** + * Class for updating views of a vector of sufficient statistics, + * in order to compute impurity from a sample. + * Note: Instances of this class do not hold the data; they operate on views of the data. + * @param numClasses Number of classes for label. + */ +private[tree] class GiniAggregator(numClasses: Int) + extends ImpurityAggregator(numClasses) with Serializable { + + /** + * Update stats for one (node, feature, bin) with the given label. + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for this (node, feature, bin). + */ + def update(allStats: Array[Double], offset: Int, label: Double): Unit = { + if (label >= statsSize) { + throw new IllegalArgumentException(s"GiniAggregator given label $label" + + s" but requires label < numClasses (= $statsSize).") + } + allStats(offset + label.toInt) += 1 + } + + /** + * Get an [[ImpurityCalculator]] for a (node, feature, bin). + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for this (node, feature, bin). + */ + def getCalculator(allStats: Array[Double], offset: Int): GiniCalculator = { + new GiniCalculator(allStats.view(offset, offset + statsSize).toArray) + } + +} + +/** + * Stores statistics for one (node, feature, bin) for calculating impurity. + * Unlike [[GiniAggregator]], this class stores its own data and is for a specific + * (node, feature, bin). + * @param stats Array of sufficient statistics for a (node, feature, bin). + */ +private[tree] class GiniCalculator(stats: Array[Double]) extends ImpurityCalculator(stats) { + + /** + * Make a deep copy of this [[ImpurityCalculator]]. + */ + def copy: GiniCalculator = new GiniCalculator(stats.clone()) + + /** + * Calculate the impurity from the stored sufficient statistics. + */ + def calculate(): Double = Gini.calculate(stats, stats.sum) + + /** + * Number of data points accounted for in the sufficient statistics. + */ + def count: Long = stats.sum.toLong + + /** + * Prediction which should be made based on the sufficient statistics. + */ + def predict: Double = if (count == 0) { + 0 + } else { + indexOfLargestArrayElement(stats) + } + + /** + * Probability of the label given by [[predict]]. + */ + override def prob(label: Double): Double = { + val lbl = label.toInt + require(lbl < stats.length, + s"GiniCalculator.prob given invalid label: $lbl (should be < ${stats.length}") + val cnt = count + if (cnt == 0) { + 0 + } else { + stats(lbl) / cnt + } + } + + override def toString: String = s"GiniCalculator(stats = [${stats.mkString(", ")}])" + +} diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Impurity.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Impurity.scala index 92b0c7b4a6..5a047d6cb5 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Impurity.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Impurity.scala @@ -22,6 +22,9 @@ import org.apache.spark.annotation.{DeveloperApi, Experimental} /** * :: Experimental :: * Trait for calculating information gain. + * This trait is used for + * (a) setting the impurity parameter in [[org.apache.spark.mllib.tree.configuration.Strategy]] + * (b) calculating impurity values from sufficient statistics. */ @Experimental trait Impurity extends Serializable { @@ -47,3 +50,127 @@ trait Impurity extends Serializable { @DeveloperApi def calculate(count: Double, sum: Double, sumSquares: Double): Double } + +/** + * Interface for updating views of a vector of sufficient statistics, + * in order to compute impurity from a sample. + * Note: Instances of this class do not hold the data; they operate on views of the data. + * @param statsSize Length of the vector of sufficient statistics for one bin. + */ +private[tree] abstract class ImpurityAggregator(val statsSize: Int) extends Serializable { + + /** + * Merge the stats from one bin into another. + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for (node, feature, bin) which is modified by the merge. + * @param otherOffset Start index of stats for (node, feature, other bin) which is not modified. + */ + def merge(allStats: Array[Double], offset: Int, otherOffset: Int): Unit = { + var i = 0 + while (i < statsSize) { + allStats(offset + i) += allStats(otherOffset + i) + i += 1 + } + } + + /** + * Update stats for one (node, feature, bin) with the given label. + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for this (node, feature, bin). + */ + def update(allStats: Array[Double], offset: Int, label: Double): Unit + + /** + * Get an [[ImpurityCalculator]] for a (node, feature, bin). + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for this (node, feature, bin). + */ + def getCalculator(allStats: Array[Double], offset: Int): ImpurityCalculator + +} + +/** + * Stores statistics for one (node, feature, bin) for calculating impurity. + * Unlike [[ImpurityAggregator]], this class stores its own data and is for a specific + * (node, feature, bin). + * @param stats Array of sufficient statistics for a (node, feature, bin). + */ +private[tree] abstract class ImpurityCalculator(val stats: Array[Double]) { + + /** + * Make a deep copy of this [[ImpurityCalculator]]. + */ + def copy: ImpurityCalculator + + /** + * Calculate the impurity from the stored sufficient statistics. + */ + def calculate(): Double + + /** + * Add the stats from another calculator into this one, modifying and returning this calculator. + */ + def add(other: ImpurityCalculator): ImpurityCalculator = { + require(stats.size == other.stats.size, + s"Two ImpurityCalculator instances cannot be added with different counts sizes." + + s" Sizes are ${stats.size} and ${other.stats.size}.") + var i = 0 + while (i < other.stats.size) { + stats(i) += other.stats(i) + i += 1 + } + this + } + + /** + * Subtract the stats from another calculator from this one, modifying and returning this + * calculator. + */ + def subtract(other: ImpurityCalculator): ImpurityCalculator = { + require(stats.size == other.stats.size, + s"Two ImpurityCalculator instances cannot be subtracted with different counts sizes." + + s" Sizes are ${stats.size} and ${other.stats.size}.") + var i = 0 + while (i < other.stats.size) { + stats(i) -= other.stats(i) + i += 1 + } + this + } + + /** + * Number of data points accounted for in the sufficient statistics. + */ + def count: Long + + /** + * Prediction which should be made based on the sufficient statistics. + */ + def predict: Double + + /** + * Probability of the label given by [[predict]], or -1 if no probability is available. + */ + def prob(label: Double): Double = -1 + + /** + * Return the index of the largest array element. + * Fails if the array is empty. + */ + protected def indexOfLargestArrayElement(array: Array[Double]): Int = { + val result = array.foldLeft(-1, Double.MinValue, 0) { + case ((maxIndex, maxValue, currentIndex), currentValue) => + if (currentValue > maxValue) { + (currentIndex, currentValue, currentIndex + 1) + } else { + (maxIndex, maxValue, currentIndex + 1) + } + } + if (result._1 < 0) { + throw new RuntimeException("ImpurityCalculator internal error:" + + " indexOfLargestArrayElement failed") + } + result._1 + } + +} diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Variance.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Variance.scala index f7d99a40eb..e9ccecb1b8 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Variance.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Variance.scala @@ -61,3 +61,75 @@ object Variance extends Impurity { def instance = this } + +/** + * Class for updating views of a vector of sufficient statistics, + * in order to compute impurity from a sample. + * Note: Instances of this class do not hold the data; they operate on views of the data. + */ +private[tree] class VarianceAggregator() + extends ImpurityAggregator(statsSize = 3) with Serializable { + + /** + * Update stats for one (node, feature, bin) with the given label. + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for this (node, feature, bin). + */ + def update(allStats: Array[Double], offset: Int, label: Double): Unit = { + allStats(offset) += 1 + allStats(offset + 1) += label + allStats(offset + 2) += label * label + } + + /** + * Get an [[ImpurityCalculator]] for a (node, feature, bin). + * @param allStats Flat stats array, with stats for this (node, feature, bin) contiguous. + * @param offset Start index of stats for this (node, feature, bin). + */ + def getCalculator(allStats: Array[Double], offset: Int): VarianceCalculator = { + new VarianceCalculator(allStats.view(offset, offset + statsSize).toArray) + } + +} + +/** + * Stores statistics for one (node, feature, bin) for calculating impurity. + * Unlike [[GiniAggregator]], this class stores its own data and is for a specific + * (node, feature, bin). + * @param stats Array of sufficient statistics for a (node, feature, bin). + */ +private[tree] class VarianceCalculator(stats: Array[Double]) extends ImpurityCalculator(stats) { + + require(stats.size == 3, + s"VarianceCalculator requires sufficient statistics array stats to be of length 3," + + s" but was given array of length ${stats.size}.") + + /** + * Make a deep copy of this [[ImpurityCalculator]]. + */ + def copy: VarianceCalculator = new VarianceCalculator(stats.clone()) + + /** + * Calculate the impurity from the stored sufficient statistics. + */ + def calculate(): Double = Variance.calculate(stats(0), stats(1), stats(2)) + + /** + * Number of data points accounted for in the sufficient statistics. + */ + def count: Long = stats(0).toLong + + /** + * Prediction which should be made based on the sufficient statistics. + */ + def predict: Double = if (count == 0) { + 0 + } else { + stats(1) / count + } + + override def toString: String = { + s"VarianceAggregator(cnt = ${stats(0)}, sum = ${stats(1)}, sum2 = ${stats(2)})" + } + +} diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/model/Bin.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/model/Bin.scala index af35d88f71..0cad473782 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/model/Bin.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/model/Bin.scala @@ -20,7 +20,7 @@ package org.apache.spark.mllib.tree.model import org.apache.spark.mllib.tree.configuration.FeatureType._ /** - * Used for "binning" the features bins for faster best split calculation. + * Used for "binning" the feature values for faster best split calculation. * * For a continuous feature, the bin is determined by a low and a high split, * where an example with featureValue falls into the bin s.t. @@ -30,13 +30,16 @@ import org.apache.spark.mllib.tree.configuration.FeatureType._ * bins, splits, and feature values. The bin is determined by category/feature value. * However, the bins are not necessarily ordered by feature value; * they are ordered using impurity. + * * For unordered categorical features, there is a 1-1 correspondence between bins, splits, * where bins and splits correspond to subsets of feature values (in highSplit.categories). + * An unordered feature with k categories uses (1 << k - 1) - 1 bins, corresponding to all + * partitionings of categories into 2 disjoint, non-empty sets. * * @param lowSplit signifying the lower threshold for the continuous feature to be * accepted in the bin * @param highSplit signifying the upper threshold for the continuous feature to be - * accepted in the bin + * accepted in the bin * @param featureType type of feature -- categorical or continuous * @param category categorical label value accepted in the bin for ordered features */ diff --git a/mllib/src/main/scala/org/apache/spark/mllib/tree/model/Node.scala b/mllib/src/main/scala/org/apache/spark/mllib/tree/model/Node.scala index 0eee626278..5b8a4cbed2 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/tree/model/Node.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/tree/model/Node.scala @@ -24,8 +24,13 @@ import org.apache.spark.mllib.linalg.Vector /** * :: DeveloperApi :: - * Node in a decision tree - * @param id integer node id + * Node in a decision tree. + * + * About node indexing: + * Nodes are indexed from 1. Node 1 is the root; nodes 2, 3 are the left, right children. + * Node index 0 is not used. + * + * @param id integer node id, from 1 * @param predict predicted value at the node * @param isLeaf whether the leaf is a node * @param split split to calculate left and right nodes @@ -51,17 +56,13 @@ class Node ( * @param nodes array of nodes */ def build(nodes: Array[Node]): Unit = { - - logDebug("building node " + id + " at level " + - (scala.math.log(id + 1)/scala.math.log(2)).toInt ) + logDebug("building node " + id + " at level " + Node.indexToLevel(id)) logDebug("id = " + id + ", split = " + split) logDebug("stats = " + stats) logDebug("predict = " + predict) if (!isLeaf) { - val leftNodeIndex = id * 2 + 1 - val rightNodeIndex = id * 2 + 2 - leftNode = Some(nodes(leftNodeIndex)) - rightNode = Some(nodes(rightNodeIndex)) + leftNode = Some(nodes(Node.leftChildIndex(id))) + rightNode = Some(nodes(Node.rightChildIndex(id))) leftNode.get.build(nodes) rightNode.get.build(nodes) } @@ -96,24 +97,20 @@ class Node ( * Get the number of nodes in tree below this node, including leaf nodes. * E.g., if this is a leaf, returns 0. If both children are leaves, returns 2. */ - private[tree] def numDescendants: Int = { - if (isLeaf) { - 0 - } else { - 2 + leftNode.get.numDescendants + rightNode.get.numDescendants - } + private[tree] def numDescendants: Int = if (isLeaf) { + 0 + } else { + 2 + leftNode.get.numDescendants + rightNode.get.numDescendants } /** * Get depth of tree from this node. * E.g.: Depth 0 means this is a leaf node. */ - private[tree] def subtreeDepth: Int = { - if (isLeaf) { - 0 - } else { - 1 + math.max(leftNode.get.subtreeDepth, rightNode.get.subtreeDepth) - } + private[tree] def subtreeDepth: Int = if (isLeaf) { + 0 + } else { + 1 + math.max(leftNode.get.subtreeDepth, rightNode.get.subtreeDepth) } /** @@ -148,3 +145,49 @@ class Node ( } } + +private[tree] object Node { + + /** + * Return the index of the left child of this node. + */ + def leftChildIndex(nodeIndex: Int): Int = nodeIndex << 1 + + /** + * Return the index of the right child of this node. + */ + def rightChildIndex(nodeIndex: Int): Int = (nodeIndex << 1) + 1 + + /** + * Get the parent index of the given node, or 0 if it is the root. + */ + def parentIndex(nodeIndex: Int): Int = nodeIndex >> 1 + + /** + * Return the level of a tree which the given node is in. + */ + def indexToLevel(nodeIndex: Int): Int = if (nodeIndex == 0) { + throw new IllegalArgumentException(s"0 is not a valid node index.") + } else { + java.lang.Integer.numberOfTrailingZeros(java.lang.Integer.highestOneBit(nodeIndex)) + } + + /** + * Returns true if this is a left child. + * Note: Returns false for the root. + */ + def isLeftChild(nodeIndex: Int): Boolean = nodeIndex > 1 && nodeIndex % 2 == 0 + + /** + * Return the maximum number of nodes which can be in the given level of the tree. + * @param level Level of tree (0 = root). + */ + def maxNodesInLevel(level: Int): Int = 1 << level + + /** + * Return the index of the first node in the given level. + * @param level Level of tree (0 = root). + */ + def startIndexInLevel(level: Int): Int = 1 << level + +} |