Spark在mllib中实现了tree相关的算法,决策树DT(DecisionTree),随机森林RF(RandomForest),GBDT(Gradient Boosting Decision Tree),其基础都是RF,DT是RF一棵树时的情况,而GBDT则是循环构建DT,GBDT与DT的代码是非常简单明了的,本文将分成五部分分别对RF的源码进行分析,介绍spark在实现过程中使用的一些技巧。
1. 决策树与随机森林
首先对决策树和随机森林进行简单的回顾。
1.1. 决策树
在决策树的训练中,如上图所示,就是从根节点开始,不断的分裂,直到触发截止条件,在节点的分裂过程中要解决的问题其实就2个
分裂点:一般就是遍历所有特征的所有特征值,选取impurity最大的分成左右孩子节点,impurity的选取有信息熵(分类),最小均方差(回归)等方法
预测值:一般取当前最多的class(分类)或者取均值(回归)
1.2. 随机森林
随机森林就是构建多棵决策树投票,在构建多棵树过程中,引入随机性,一般体现在两个方面,一是每棵树使用的样本进行随机抽样,分为有放回和无放回抽样。二是对每棵树使用的特征集进行抽样,使用部分特征训练。
在训练过程中,如果单机内存能放下所有样本,可以用多线程同时训练多棵树,树之间的训练互不影响。
2. spark RF优化策略
spark在实现RF时,使用了一些优化技巧,提高训练效率。
2.1. 逐层训练
当样本量过大,单机无法容纳时,只能采用分布式的训练方法,数据是在集群中的多台机器存放,如果按照单机的方法,每棵树完全独立访问样本数据,则样本数据的访问次数为数的个数k*每棵树的节点数N,相当于深度遍历。在spark的实现中,因为数据存放在不同的机器上,频繁的访问数据效率非常低,因此采用广度遍历的方法,每次构造所有树的一层,例如如果要训练10棵树,第一次构造所有树的第一层根节点,第二次构造所有深度为2的节点,以此类推,这样访问数据的次数降为树的最大深度,大大减少了机器之间的通信,提高训练效率。
2.2. 样本抽样
当样本存在连续特征时,其可能的取值可能是无限的,存储其可能出现的值占用较大空间,因此spark对样本进行了抽样,抽样数量
val requiredSamples = math.max(metadata.maxBins * metadata.maxBins, 10000)
最少抽样1万条,当然这样会降低模型精度。
2.3. 特征装箱
其实没什么神秘的,每个离散特征值(对于连续特征,先离散化)称为一个Split,上下限[lowSplit, highSplit]组成一个bin,也就是特征装箱,默认的maxBins是32。对于连续特征,离散化时的bin的个数就是maxBins,采用等频离散化;对于有序的离散特征,bin的个数是特征值个数+1;对于无序离散特征,bin的个数是2^(M-1)-1,M是特征值个数
3. 源码分析
我们从官方给出的分类demo开始,逐层分析其实现
3.1. 训练数据的解析
主要是LabelPoint的构造,官方demo中要求训练数据是LibSVM格式的
parsed.map { case (label, indices, values) =>
LabeledPoint(label, Vectors.sparse(d, indices, values))
}
可以看到LabelPoint有两个成员,第一个是样本label,第二个是稀疏向量SparseVector,d是其size,在这里其实是特征数,indices是实际非0特征的index,values里面是实际的特征值,这里需要注意的是,SVN格式的特征index是从0开始的,这里进行了-1,变成从0开始了。
3.2. demo中训练参数说明
官方demo中只设置了部分参数
val model = RandomForest.trainClassifier(trainingData, numClasses, categoricalFeaturesInfo,
numTrees, featureSubsetStrategy, impurity, maxDepth, maxBins)
categoricalFeaturesInfo:Map[Int, Int],key是特征的index,value为特征值的个数(或者说几种),这里值得注意的是,因为LabelPoint中进行了index-1的变换,这个里面的key也需要-1(参见后面metadata的numBins的计算)。例如性别这个特征在样本中的index为1,特征值男/女两种,则0->2
featureSubsetStrategy:特征子集的抽取方法,支持”auto”, “all”, “sqrt”, “log2”, “onethird”
impurity:不纯度,其实就是节点分裂时的衡量准则,例如信息熵,均方差等,这里支持三种,gini(基尼指数),entripy(信息熵),variance(均方差)
maxDepth:树的最大深度
maxBins:最大装箱数,或者说是特征的最大可能切分数+1。这个值必须大于等于最大的离散特征值数
3.3. 参数封装
spark根据用户提供的参数值,进行实际训练参数的计算,并且将这些参数封装成类,方便传递。
3.3.1. Strategy
class Strategy @Since("1.3.0") (
@Since("1.0.0") @BeanProperty var algo: Algo,
@Since("1.0.0") @BeanProperty var impurity: Impurity,
@Since("1.0.0") @BeanProperty var maxDepth: Int,
@Since("1.2.0") @BeanProperty var numClasses: Int = 2,
@Since("1.0.0") @BeanProperty var maxBins: Int = 32,
@Since("1.0.0") @BeanProperty var quantileCalculationStrategy: QuantileStrategy = Sort,
@Since("1.0.0") @BeanProperty var categoricalFeaturesInfo: Map[Int, Int] = Map[Int, Int](),
@Since("1.2.0") @BeanProperty var minInstancesPerNode: Int = 1,
@Since("1.2.0") @BeanProperty var minInfoGain: Double = 0.0,
@Since("1.0.0") @BeanProperty var maxMemoryInMB: Int = 256,
@Since("1.2.0") @BeanProperty var subsamplingRate: Double = 1,
@Since("1.2.0") @BeanProperty var useNodeIdCache: Boolean = false,
@Since("1.2.0") @BeanProperty var checkpointInterval: Int = 10)
algo:classification/regression
quantileCalculationStrategy:分位点(Split)策略,目前只支持Sort,对于连续型特征值,先把特征值进行排序,然后按次序取分位点。从代码中可以看到原来可能打算实现的MinMax和ApproxHist目前没有实现。
minInstancesPerNode:每个树节点中最小的样本数,低于将不再对节点进行分裂,默认为1,可作为提前截止条件
minInfoGain:最小增益,节点分裂后的增益如果小于它,将不再进行分裂,可作为提前截止条件
subsamplingRate:样本抽样率,默认为1,每棵树都使用全部样本
isMulticlassClassification:是否是多分类,判断条件为Classification 并且类别>2
isMulticlassWithCategoricalFeatures:是否是带类别特征的多分类,判断条件再上面的基础上加categoricalFeaturesInfo的size大于0
3.3.2. metadata
在buildMetadata中根据strategy计算得到DecisionTreeMetadata的参数。
class DecisionTreeMetadata(
val numFeatures: Int,
val numExamples: Long,
val numClasses: Int,
val maxBins: Int,
val featureArity: Map[Int, Int],
val unorderedFeatures: Set[Int],
val numBins: Array[Int],
val impurity: Impurity,
val quantileStrategy: QuantileStrategy,
val maxDepth: Int,
val minInstancesPerNode: Int,
val minInfoGain: Double,
val numTrees: Int,
val numFeaturesPerNode: Int)
部分参数同Strategy,对额外参数和区别说明
numClasses:如为Regression,设为0
maxPossibleBins:取maxBins和样本数量中较小的;必须大于categoricalFeaturesInfo中的最大的离散特征值数
numBins:所有特征及其特征值数,Int数组,维数是特征数,默认大小是maxPossibleBins。对于连续特征,其值就是默认值maxPossibleBins。对于离散特征,如为二分类或回归,此处将categoricalFeaturesInfo中的key特征index作为数组index,value特征个数写入数组中(这里有疑问,SVM格式的index是从1开始的,因此对numBins的index应该是categoricalFeaturesInfo的key-1,这里没有-1,当最大值等于maxBins的时候访问数组会抛异常);如果是多分类,先计算其当做当UnorderedFeature(无序的离散特征)的bin,如果个数小于等于maxPossibleBins,会被当成UnorderedFeature,否则被当成orderedFeatures(为了防止计算指数溢出,实际是把maxPossibleBins取log与特征数比较),因为UnorderedFeature的bin是比较大,这里限制了其特征值不能太多,这里仅仅根据特征值的特殊决定是否是ordered,不太好。每个split要将所有特征值分成2部分,bin的数量也就是2*split,因此bin的个数是2*(2^(M-1)-1)
numFeaturesPerNode:由featureSubsetStrategy决定,如果为“auto”,且为单棵树,则使用全部特征;如为多棵树,分类则是sqrt,回归为1/3;也可以自己指定,支持”all”, “sqrt”, “log2”, “onethird”。
如果仅对RF的使用感兴趣,了解上述训练参数也就可以了,后面的文章将对其训练代码进行分析。
4. 特征处理
这部分主要在DecisionTree.scala的findSplitsBins函数,将所有特征封装成Split,然后装箱Bin。首先对split和bin的结构进行说明
4.1. 数据结构
4.1.1. Split
class Split(
@Since("1.0.0") feature: Int,
@Since("1.0.0") threshold: Double,
@Since("1.0.0") featureType: FeatureType,
@Since("1.0.0") categories: List[Double])
feature:特征id
threshold:阈值
featureType:连续特征(Continuous)/离散特征(Categorical)
categories:离散特征值数组,离散特征使用。放着此split中所有特征值
4.1.2. Bin
class Bin(
lowSplit: Split,
highSplit: Split,
featureType: FeatureType,
category: Double)
lowSplit/highSplit:上下界
featureType:连续特征(Continuous)/离散特征(Categorical)
category:离散特征的特征值
4.2. 连续特征处理
4.2.1. 抽样
val continuousFeatures = Range(0, numFeatures).filter(metadata.isContinuous)
val sampledInput = if (continuousFeatures.nonEmpty) {
// 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())
} else {
input.sparkContext.emptyRDD[LabeledPoint]
}
首先筛选出连续特征集,然后计算抽样数量,抽样比例,然后无放回样本抽样;如果没有连续特征,则为空RDD
4.2.2. 计算Split
metadata.quantileStrategy match {
case Sort =>
findSplitsBinsBySorting(sampledInput, metadata, continuousFeatures)
case MinMax =>
throw new UnsupportedOperationException("minmax not supported yet.")
case ApproxHist =>
throw new UnsupportedOperationException("approximate histogram not supported yet.")
}
分位点策略,这里只实现了Sort这一种,前文有说明,下面的计算在findSplitsBinsBySorting函数中,入参是抽样样本集,metadata和连续特征集(里面是特征id,从0开始,见LabelPoint的构造)
val continuousSplits = {
// reduce the parallelism for split computations when there are less
// continuous features than input partitions. this prevents tasks from
// being spun up that will definitely do no work.
val numPartitions = math.min(continuousFeatures.length,input.partitions.length)
input.flatMap(point => continuousFeatures.map(idx => (idx,point.features(idx))))
.groupByKey(numPartitions)
.map { case (k, v) => findSplits(k, v) }
.collectAsMap()
}
特征id为key,value是样本对应的该特征下的所有特征值,传给findSplits函数,其中又调用了findSplitsForContinuousFeature函数获得连续特征的Split,入参为样本,metadata和特征id
def findSplitsForContinuousFeature(
featureSamples: Array[Double],
metadata: DecisionTreeMetadata,
featureIndex: Int): Array[Double] = {
require(metadata.isContinuous(featureIndex),
"findSplitsForContinuousFeature can only be used to find splits for a continuous feature.")
val splits = {
//连续特征的split是numBins-1
val numSplits = metadata.numSplits(featureIndex)
//统计所有特征值其出现的次数
// get count for each distinct value
val valueCountMap = featureSamples.foldLeft(Map.empty[Double, Int]) { (m, x) =>
m + ((x, m.getOrElse(x, 0) + 1))
}
//按特征值排序
// sort distinct values
val valueCounts = valueCountMap.toSeq.sortBy(_._1).toArray
// if possible splits is not enough or just enough, just return all possible splits
val possibleSplits = valueCounts.length
if (possibleSplits <= numSplits) {
valueCounts.map(_._1)
} else {
//等频离散化
// stride between splits
val stride: Double = featureSamples.length.toDouble / (numSplits + 1)
logDebug("stride = " + stride)
// iterate `valueCount` to find splits
val splitsBuilder = Array.newBuilder[Double]
var index = 1
// currentCount: sum of counts of values that have been visited
var currentCount = valueCounts(0)._2
// targetCount: target value for `currentCount`.
// If `currentCount` is closest value to `targetCount`,
// then current value is a split threshold.
// After finding a split threshold, `targetCount` is added by stride.
var targetCount = stride
while (index < valueCounts.length) {
val previousCount = currentCount
currentCount += valueCounts(index)._2
val previousGap = math.abs(previousCount - targetCount)
val currentGap = math.abs(currentCount - targetCount)
// If adding count of current value to currentCount
// makes the gap between currentCount and targetCount smaller,
// previous value is a split threshold.
//每次步进targetCount个样本,取上一个特征值与下一个特征值gap较小的
if (previousGap < currentGap) {
splitsBuilder += valueCounts(index - 1)._1
targetCount += stride
}
index += 1
}
splitsBuilder.result()
}
}
// TODO: Do not fail; just ignore the useless feature.
assert(splits.length > 0,
s"DecisionTree could not handle feature $featureIndex since it had only 1 unique value." +
" Please remove this feature and then try again.")
// the split metadata must be updated on the driver
splits
}
在构造split的过程中,如果统计到的值的个数possibleSplits 还不如你设置的numSplits多,那么所有的值都作为分割点;否则,用等频分隔法,首先计算分隔步长stride,然后再循环中每次累加到targetCount中,作为理想分割点,但是理想分割点可能会包含的特征值过多,想取一个里理想分割点尽量近的特征值,例如,理想分割点是100,落在特征值fcfc里,但是当前特征值里面有30个样本,而前一个特征值fpfp只有5个样本,因此我们如果取fcfc作为split,则当前区间实际多25个样本,如果取fpfp,则少5个样本,显然取fpfp更为合理。
具体到代码实现,在if判断里步进stride个样本,累加在targetCount中。while循环逐次把每个特征值的个数加到currentCount里,计算前一次previousCount和这次currentCount到targetCount的距离,有3种情况,一种是pre和cur都在target左边,肯定是cur小,继续循环,进入第二种情况;第二种一左一右,如果pre小,肯定是pre是最好的分割点,如果cur还是小,继续循环步进,进入第三种情况;第三种就是都在右边,显然是pre小。因此if的判断条件pre<curpre<cur,只要满足肯定就是split。整体下来的效果就能找到离target最近的一个特征值。
findSplits函数使用本函数得到的离散化点作为threshold,构造Split
val splits = {
val featureSplits = findSplitsForContinuousFeature(
featureSamples.toArray,
metadata,
featureIndex)
logDebug(s"featureIndex = $featureIndex, numSplits = ${featureSplits.length}")
featureSplits.map(threshold => new Split(featureIndex, threshold, Continuous, Nil))
}
这样就得到了连续特征所有的Split
4.2.3. 计算bin
得到splits后,即可类似滑窗得到bin的上下界,构造bins
val bins = {
val lowSplit = new DummyLowSplit(featureIndex, Continuous)
val highSplit = new DummyHighSplit(featureIndex, Continuous)
// tack the dummy splits on either side of the computed splits
val allSplits = lowSplit +: splits.toSeq :+ highSplit
// slide across the split points pairwise to allocate the bins
allSplits.sliding(2).map {
case Seq(left, right) => new Bin(left, right, Continuous, Double.MinValue)
}.toArray
}
在计算splits的时候,个数是bin的个数减1,这里加上第一个DummyLowSplit(threshold是Double.MinValue),和最后一个DummyHighSplit(threshold是Double.MaxValue)构造的bin,恰好个数是numBins中的个数
4.3. 离散特征
bin的主要作用其实就是用来做连续特征离散化,离散特征是用不着的。
对有序离散特征而言,其split直接用特征值表征,因此这里的splits和bins都是空的Array。
对于无序离散特征而言,其split是特征值的组合,不是简单的上下界比较关系,bin是空Array,而split需要计算。
4.3.1. split
// Unordered features
// 2^(maxFeatureValue - 1) - 1 combinations
val featureArity = metadata.featureArity(i)
val split = Range(0, metadata.numSplits(i)).map { splitIndex =>
val categories = extractMultiClassCategories(splitIndex + 1, featureArity)
new Split(i, Double.MinValue, Categorical, categories)
}
featureArity来自参数categoricalFeaturesInfo中设置的离散特征的特征值数。
metadata.numSplits是吧numBins中的数量/2,相当于返回了2^(M-1)-1,M是特征值数。
调用extractMultiClassCategories函数,入参是1到2^(M-1)和特征数M。
/**
* Nested method to extract list of eligible categories given an index. It extracts the
* position of ones in a binary representation of the input. If binary
* representation of an number is 01101 (13), the output list should (3.0, 2.0,
* 0.0). The maxFeatureValue depict the number of rightmost digits that will be tested for ones.
*/
def extractMultiClassCategories(
input: Int,
maxFeatureValue: Int): List[Double] = {
var categories = List[Double]()
var j = 0
var bitShiftedInput = input
while (j < maxFeatureValue) {
if (bitShiftedInput % 2 != 0) {
// updating the list of categories.
categories = j.toDouble :: categories
}
// Right shift by one
bitShiftedInput = bitShiftedInput >> 1
j += 1
}
categories
}
如注释所述,这个函数返回给定的input的二进制表示中1的index,这里实际返回的是特征的组合,之前文章介绍过的《组合数》。
5. 样本处理
将输入样本LabelPoint与上述特征进一步封装,方便后面进行分区统计。
5.1. TreePoint
构造TreePoint的过程,是一系列函数的调用链,我们逐层分析。
val treeInput = TreePoint.convertToTreeRDD(retaggedInput, bins, metadata)
RandomForest.scala中将输入转化成TreePoint的rdd,调用convertToTreeRDD函数
def convertToTreeRDD(
input: RDD[LabeledPoint],
bins: Array[Array[Bin]],
metadata: DecisionTreeMetadata): RDD[TreePoint] = {
// Construct arrays for featureArity for efficiency in the inner loop.
val featureArity: Array[Int] = new Array[Int](metadata.numFeatures)
var featureIndex = 0
while (featureIndex < metadata.numFeatures) {
featureArity(featureIndex) = metadata.featureArity.getOrElse(featureIndex, 0)
featureIndex += 1
}
input.map { x =>
TreePoint.labeledPointToTreePoint(x, bins, featureArity)
}
}
convertToTreeRDD函数的入参input是所有样本,bins是二维数组,第一维是特征,第二维是特征的Bin数组。函数首先计算每个特征的特征数量,放在featureArity中,如果是连续特征,设为0。对每个样本调用labeledPointToTreePoint函数,构造TreePoint。
private def labeledPointToTreePoint(
labeledPoint: LabeledPoint,
bins: Array[Array[Bin]],
featureArity: Array[Int]): TreePoint = {
val numFeatures = labeledPoint.features.size
val arr = new Array[Int](numFeatures)
var featureIndex = 0
while (featureIndex < numFeatures) {
arr(featureIndex) = findBin(featureIndex, labeledPoint, featureArity(featureIndex),
bins)
featureIndex += 1
}
new TreePoint(labeledPoint.label, arr)
}
labeledPointToTreePoint计算每个样本的所有特征对应的特征值属于哪个bin,放在在arr数组中;如果是连续特征,存放的实际是binIndex,或者说是第几个bin;如果是离散特征,直接featureValue.toInt,这其实暗示着,对有序离散值,其编码只能是[0,featureArity - 1],闭区间,其后的部分逻辑也依赖于这个假设。这部分是在findBin函数中完成的,这里不再赘述。
我们在这里把TreePoint的成员再罗列一下,方便查阅
class TreePoint(val label: Double, val binnedFeatures: Array[Int])
这里是把每个样本从LabelPoint转换成TreePoint,label就是样本label,binnedFeatures就是上述的arr数组。
5.2. BaggedPoint
同理构造BaggedPoint的过程,也是一系列函数的调用链,我们逐层分析。
val withReplacement = if (numTrees > 1) true else false
val baggedInput = BaggedPoint.convertToBaggedRDD(treeInput,
strategy.subsamplingRate, numTrees,
withReplacement, seed).persist(StorageLevel.MEMORY_AND_DISK)
这里同时对样本进行了抽样,如果树个数大于1,就有放回抽样,否则无放回抽样,调用convertToTreeRDD函数将TreePoint转化成BaggedPoint的rdd
/**
* Convert an input dataset into its BaggedPoint representation,
* choosing subsamplingRate counts for each instance.
* Each subsamplingRate has the same number of instances as the original dataset,
* and is created by subsampling without replacement.
* @param input Input dataset.
* @param subsamplingRate Fraction of the training data used for learning decision tree.
* @param numSubsamples Number of subsamples of this RDD to take.
* @param withReplacement Sampling with/without replacement.
* @param seed Random seed.
* @return BaggedPoint dataset representation.
*/
def convertToBaggedRDD[Datum] (
input: RDD[Datum],
subsamplingRate: Double,
numSubsamples: Int,
withReplacement: Boolean,
seed: Long = Utils.random.nextLong()): RDD[BaggedPoint[Datum]] = {
if (withReplacement) {
convertToBaggedRDDSamplingWithReplacement(input, subsamplingRate, numSubsamples, seed)
} else {
if (numSubsamples == 1 && subsamplingRate == 1.0) {
convertToBaggedRDDWithoutSampling(input)
} else {
convertToBaggedRDDSamplingWithoutReplacement(input, subsamplingRate, numSubsamples, seed)
}
}
}
根据有放回还是无放回,或者不抽样分别调用相应函数。无放回抽样
def convertToBaggedRDDSamplingWithoutReplacement[Datum] (
input: RDD[Datum],
subsamplingRate: Double,
numSubsamples: Int,
seed: Long): RDD[BaggedPoint[Datum]] = {
//对每个partition独立抽样
input.mapPartitionsWithIndex { (partitionIndex, instances) =>
// Use random seed = seed + partitionIndex + 1 to make generation reproducible.
val rng = new XORShiftRandom
rng.setSeed(seed + partitionIndex + 1)
instances.map { instance =>
//对每条样本进行numSubsamples(实际是树的个数)次抽样,
//一次将本条样本在所有树中是否会被抽取都获得,牺牲空间减少访问数据次数
val subsampleWeights = new Array[Double](numSubsamples)
var subsampleIndex = 0
while (subsampleIndex < numSubsamples) {
val x = rng.nextDouble()
//无放回抽样,只需要决定本样本是否被抽取,被抽取就是1,没有就是0
subsampleWeights(subsampleIndex) = {
if (x < subsamplingRate) 1.0 else 0.0
}
subsampleIndex += 1
}
new BaggedPoint(instance, subsampleWeights)
}
}
}
有放回抽样
def convertToBaggedRDDSamplingWithReplacement[Datum] (
input: RDD[Datum],
subsample: Double,
numSubsamples: Int,
seed: Long): RDD[BaggedPoint[Datum]] = {
input.mapPartitionsWithIndex { (partitionIndex, instances) =>
// Use random seed = seed + partitionIndex + 1 to make generation reproducible.
val poisson = new PoissonDistribution(subsample)
poisson.reseedRandomGenerator(seed + partitionIndex + 1)
instances.map { instance =>
val subsampleWeights = new Array[Double](numSubsamples)
var subsampleIndex = 0
while (subsampleIndex < numSubsamples) {
//与无放回抽样对比,这里用泊松抽样返回的是样本被抽取的次数,
//可能大于1,而无放回是0/1,也可认为是被抽取的次数
subsampleWeights(subsampleIndex) = poisson.sample()
subsampleIndex += 1
}
new BaggedPoint(instance, subsampleWeights)
}
}
}
不抽样,或者说抽样率为1
def convertToBaggedRDDWithoutSampling[Datum] (
input: RDD[Datum]): RDD[BaggedPoint[Datum]] = {
input.map(datum => new BaggedPoint(datum, Array(1.0)))
}
这里再啰嗦的罗列下BaggedPoint
class BaggedPoint[Datum](
val datum: Datum,
val subsampleWeights: Array[Double])
datum是TreePoint,subsampleWeights是数组,维数等于numberTrees,每个值是样本在每棵树中被抽取的次数
至此,Random Forest的初始化工作已经完成
timer.stop("init")
6. 随机森林训练
6.1. 数据结构
6.1.1. Node
树中的每个节点是一个Node结构
class Node @Since("1.2.0") (
@Since("1.0.0") val id: Int,
@Since("1.0.0") var predict: Predict,
@Since("1.2.0") var impurity: Double,
@Since("1.0.0") var isLeaf: Boolean,
@Since("1.0.0") var split: Option[Split],
@Since("1.0.0") var leftNode: Option[Node],
@Since("1.0.0") var rightNode: Option[Node],
@Since("1.0.0") var stats: Option[InformationGainStats])
emptyNode,只初始化nodeIndex,其他都是默认值
def emptyNode(nodeIndex: Int): Node =
new Node(nodeIndex, new Predict(Double.MinValue),
-1.0, false, None, None, None, None)
根据node的id,计算孩子节点的id
* 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
左孩子节点就是当前id * 2,右孩子是id * 2+1。
6.1.2. Entropy
6.1.2.1. Entropy
Entropy是个Object,里面最重要的是calculate函数
/**
* :: DeveloperApi ::
* information calculation for multiclass classification
* @param counts Array[Double] with counts for each label
* @param totalCount sum of counts for all labels
* @return information value, or 0 if totalCount = 0
*/
@Since("1.1.0")
@DeveloperApi
override def calculate(counts: Array[Double], totalCount: Double): Double = {
if (totalCount == 0) {
return 0
}
val numClasses = counts.length
var impurity = 0.0
var classIndex = 0
while (classIndex < numClasses) {
val classCount = counts(classIndex)
if (classCount != 0) {
val freq = classCount / totalCount
impurity -= freq * log2(freq)
}
classIndex += 1
}
impurity
}
熵的计算公式
H=E[−logpi]=−∑i=1n−pilogpi
H=E[−logpi]=−∑i=1n−pilogpi
因此这里的入参count是各class的出现的次数,先计算出现概率,然后取log累加。
6.1.2.2. EntropyAggregator
class EntropyAggregator(numClasses: Int)
extends ImpurityAggregator(numClasses)
只有一个成员变量class的个数,关键是update函数
/**
* 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, instanceWeight: Double): Unit = {
if (label >= statsSize) {
throw new IllegalArgumentException(s"EntropyAggregator given label $label" +
s" but requires label < numClasses (= $statsSize).")
}
if (label < 0) {
throw new IllegalArgumentException(s"EntropyAggregator given label $label" +
s"but requires label is non-negative.")
}
allStats(offset + label.toInt) += instanceWeight
}
offset是特征值偏移,加上label就是该class在allStats里的位置,累加出现的次数
/**
* 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)
}
截取allStats中属于该特征的split的部分数组,长度是statSize,也就是class数
6.1.2.3. EntropyCalculator
/**
* Calculate the impurity from the stored sufficient statistics.
*/
def calculate(): Double = Entropy.calculate(stats, stats.sum)
结合上面的函数可以看到,计算entropy的路径是调用Entropy的getCalculator函数,里面截取allStats中属于该split的部分,然后实际调用Entropy的calculate函数计算熵。
这里还重载了prob函数,主要是返回label的概率,例如0的统计有3个,1的统计7个,则label 0的概率就是0.3.
6.1.3. DTStatsAggregator
这里啰嗦下node分裂时需要怎样统计,这与DTStatsAggregator的设计是相关的。以使用信息熵为例,node分裂时,迭代每个特征的每个split,这个split会把样本集分成两部分,要计算entropy,需要分别统计左/右部分class的分布情况,然后计算概率,进而计算entropy,因此aggregator中statsSize等于numberclasses,同时allStats里记录了所有的统计值,实际这个统计值就是class的分布情况
class DTStatsAggregator(
val metadata: DecisionTreeMetadata,
featureSubset: Option[Array[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.
*/
private val statsSize: Int = impurityAggregator.statsSize
/**
* Number of bins for each feature. This is indexed by the feature index.
*/
private val numBins: Array[Int] = {
if (featureSubset.isDefined) {
featureSubset.get.map(metadata.numBins(_))
} else {
metadata.numBins
}
}
/**
* Offset for each feature for calculating indices into the [[allStats]] array.
*/
private val featureOffsets: Array[Int] = {
numBins.scanLeft(0)((total, nBins) => total + statsSize * nBins)
}
/**
* Total number of elements stored in this aggregator
*/
private val allStatsSize: Int = featureOffsets.last
/**
* Flat array of elements.
* Index for start of stats for a (feature, bin) is:
* index = featureOffsets(featureIndex) + binIndex * statsSize
* Note: For unordered features,
* the left child stats have binIndex in [0, numBins(featureIndex) / 2))
* and the right child stats in [numBins(featureIndex) / 2), numBins(featureIndex))
*/
private val allStats: Array[Double] = new Array[Double](allStatsSize)
每个node有一个DTStatsAggregator,构造函数接受2个参数,metadata和node使用的特征子集。其他的类成员
- impurityAggregator:目前支持Gini,Entropy和Variance,后面我们以Entropy为例,其他类似
- statsSize:每个bin需要的统计数,分类时等于numClasses,因为于每个class都需要单独统计;回归等于3,分别存着特征值个数,特征值sum,特征值平方和,为计算variance
- numBins:node所用特征对应的numBins数组元素
- featureOffsets:计算特征在allStats中的index,与每个特征的bin个数和statsSize有关,例如我们有3个特征,其bins分别为3,2,2,statsSize为2,则第一个特征需要的bin的个数是3 * 2=6,2 * 2=4,2 * 2=4,则featureOffsets为0,6,10,14,是从左到右的累计值
- allStatsSize:需要的桶的个数
- allStats:存储统计值的桶
f0,f1,f2是3个特征,f0有3个特征值(其实是binIndex)0/1/2,f1有2个0/1,f2有2个0/1,每个特征值都有statsSize个状态桶,因此共14个,个数allStatsSize=14, 比如我们想在f1的v1的c1的index,就是从featureOffsets中取得f1的特征偏移量featureOffsets(1)=6,v1的binIndex相当于是1,statsSize是2,其label是1,则桶的index=6+1*2+1=9,恰好是图中f1v1的c1的桶的index
我们对其中的关键函数进行说明
/**
* Update the stats for a given (feature, bin) for ordered features, using the given label.
*/
def update(featureIndex: Int, binIndex: Int, label: Double, instanceWeight: Double): Unit = {
//第一部分是特征偏移
//binIndex相当于特征内特征值的偏移,每个特征有statsSize个桶,因此两者相加就是这个特征值对应的桶
//例如Entropy的update函数,里面再加上label.toInt就是这个label的桶
//从这里特征偏移的计算可以看出ordered特征其特征值最好是连续的,中间无间断,并且必须从0开始
//当然如果有间断,这里相当于浪费部分空间
val i = featureOffsets(featureIndex) + binIndex * statsSize
impurityAggregator.update(allStats, i, label, instanceWeight)
}
/**
* Get an [[ImpurityCalculator]] for a given (node, feature, bin).
* @param featureOffset For ordered features, this is a pre-computed (node, feature) offset
* from [[getFeatureOffset]].
* For unordered features, this is a pre-computed
* (node, feature, left/right child) offset from
* [[getLeftRightFeatureOffsets]].
*/
def getImpurityCalculator(featureOffset: Int, binIndex: Int): ImpurityCalculator = {
//偏移的计算同上,不过这里特征偏移是入参给出的,不需要再计算
impurityAggregator.getCalculator(allStats, featureOffset + binIndex * statsSize)
}
6.2. 训练初始化
// FIFO queue of nodes to train: (treeIndex, node)
val nodeQueue = new mutable.Queue[(Int, Node)]()
val topNodes: Array[Node] = Array.fill[Node](numTrees)(Node.emptyNode(nodeIndex = 1))
Range(0, numTrees).foreach(treeIndex => nodeQueue.enqueue((treeIndex, topNodes(treeIndex))))
构造了numTrees个Node,赋默认值emptyNode,这些node将作为每棵树的root node,参与后面的训练。将这些node与treeIndex封装加入到队列nodeQueue中,后面会将所有待split的node都加入到这个队列中,依次split,直到所有node触发截止条件,也就是后面的while循环中队列为空了。
6.3. 选择待分裂node
这部分逻辑在selectNodesToSplit中,主要是从nodeQueue中取出本轮需要分裂的node,并计算node的参数。
/**
* Pull nodes off of the queue, and collect a group of nodes to be split on this iteration.
* This tracks the memory usage for aggregates and stops adding nodes when too much memory
* will be needed; this allows an adaptive number of nodes since different nodes may require
* different amounts of memory (if featureSubsetStrategy is not "all").
*
* @param nodeQueue Queue of nodes to split.
* @param maxMemoryUsage Bound on size of aggregate statistics.
* @return (nodesForGroup, treeToNodeToIndexInfo).
* nodesForGroup holds the nodes to split: treeIndex --> nodes in tree.
*
* treeToNodeToIndexInfo holds indices selected features for each node:
* treeIndex --> (global) node index --> (node index in group, feature indices).
* The (global) node index is the index in the tree; the node index in group is the
* index in [0, numNodesInGroup) of the node in this group.
* The feature indices are None if not subsampling features.
*/
private[tree] def selectNodesToSplit(
nodeQueue: mutable.Queue[(Int, Node)],
maxMemoryUsage: Long,
metadata: DecisionTreeMetadata,
rng: scala.util.Random): (Map[Int, Array[Node]], Map[Int, Map[Int, NodeIndexInfo]]) = {
// Collect some nodes to split:
// nodesForGroup(treeIndex) = nodes to split
val mutableNodesForGroup = new mutable.HashMap[Int, mutable.ArrayBuffer[Node]]()
val mutableTreeToNodeToIndexInfo =
new mutable.HashMap[Int, mutable.HashMap[Int, NodeIndexInfo]]()
var memUsage: Long = 0L
var numNodesInGroup = 0
while (nodeQueue.nonEmpty && memUsage < maxMemoryUsage) {
val (treeIndex, node) = nodeQueue.head
//用蓄水池抽样(之前的文章有介绍)对node使用的特征集抽样
// Choose subset of features for node (if subsampling).
val featureSubset: Option[Array[Int]] = if (metadata.subsamplingFeatures) {
Some(SamplingUtils.reservoirSampleAndCount(Range(0,
metadata.numFeatures).iterator, metadata.numFeaturesPerNode, rng.nextLong)._1)
} else {
None
}
// Check if enough memory remains to add this node to the group.
val nodeMemUsage = RandomForest.aggregateSizeForNode(metadata, featureSubset) * 8L
if (memUsage + nodeMemUsage <= maxMemoryUsage) {
nodeQueue.dequeue()
mutableNodesForGroup.getOrElseUpdate(treeIndex, new mutable.ArrayBuffer[Node]()) += node
mutableTreeToNodeToIndexInfo
.getOrElseUpdate(treeIndex, new mutable.HashMap[Int, NodeIndexInfo]())(node.id)
= new NodeIndexInfo(numNodesInGroup, featureSubset)
}
numNodesInGroup += 1
memUsage += nodeMemUsage
}
// Convert mutable maps to immutable ones.
val nodesForGroup: Map[Int, Array[Node]] = mutableNodesForGroup.mapValues(_.toArray).toMap
val treeToNodeToIndexInfo = mutableTreeToNodeToIndexInfo.mapValues(_.toMap).toMap
(nodesForGroup, treeToNodeToIndexInfo)
}
代码比较简单明确,受限于内存,将本次能够处理的node从nodeQueue中取出,放入nodesForGroup和treeToNodeToIndexInfo中。
是否对特征集进行抽样的条件是metadata的 numFeatures是否等于numFeaturesPerNode,这两个参数是metadata的入参,在buildMetadata时,根据featureSubsetStrateg确定,参见前文。
nodesForGroup是Map[Int, Array[Node]],其key是treeIndex,value是Node数组,其中放着该tree本次要分裂的node。
treeToNodeToIndexInfo的类型是Map[Int, Map[Int, NodeIndexInfo]],key为treeIndex,value中Map的key是node.id,这个id来自Node初始化时的第一个参数,第一轮时node的id都是1。其value为NodeIndexInfo结构,
class NodeIndexInfo(
val nodeIndexInGroup: Int,
val featureSubset: Option[Array[Int]])
第一个成员是此node在本次node选择的while循环中的index,称为groupIndex,第二个成员是特征子集。
6.4. node分裂
逻辑主要在DecisionTree.findBestSplits函数中,是RF训练最核心的部分
DecisionTree.findBestSplits(baggedInput, metadata, topNodes, nodesForGroup,
treeToNodeToIndexInfo, splits, bins, nodeQueue, timer, nodeIdCache = nodeIdCache)
6.4.1. 数据统计
数据统计分成两部分,先在各个partition上分别统计,再累积各partition成全局统计。
6.4.1.1. 取出node的特征子集
val nodeToFeatures = getNodeToFeatures(treeToNodeToIndexInfo)
val nodeToFeaturesBc = input.sparkContext.broadcast(nodeToFeatures)
取出各node的特征子集,如果不需要抽样则为None;否则返回Map[Int, Array[Int]],其实就是将之前treeToNodeToIndexInfo中的NodeIndexInfo转换为map结构,将其作为广播变量nodeToFeaturesBc。
6.4.1.2. 分区统计
一系列函数的调用链,我们逐层分析
val partitionAggregates : RDD[(Int, DTStatsAggregator)] = if (nodeIdCache.nonEmpty) {
input.zip(nodeIdCache.get.nodeIdsForInstances).mapPartitions { points =>
// Construct a nodeStatsAggregators array to hold node aggregate stats,
// each node will have a nodeStatsAggregator
val nodeStatsAggregators = Array.tabulate(numNodes) { nodeIndex =>
val featuresForNode = nodeToFeaturesBc.value.flatMap { nodeToFeatures =>
Some(nodeToFeatures(nodeIndex))
}
new DTStatsAggregator(metadata, featuresForNode)
}
// iterator all instances in current partition and update aggregate stats
points.foreach(binSeqOpWithNodeIdCache(nodeStatsAggregators, _))
// transform nodeStatsAggregators array to (nodeIndex, nodeAggregateStats) pairs,
// which can be combined with other partition using `reduceByKey`
nodeStatsAggregators.view.zipWithIndex.map(_.swap).iterator
}
} else {
input.mapPartitions { points =>
// Construct a nodeStatsAggregators array to hold node aggregate stats,
// each node will have a nodeStatsAggregator
val nodeStatsAggregators = Array.tabulate(numNodes) { nodeIndex =>
val featuresForNode = nodeToFeaturesBc.value.flatMap { nodeToFeatures =>
Some(nodeToFeatures(nodeIndex))
}
new DTStatsAggregator(metadata, featuresForNode)
}
// iterator all instances in current partition and update aggregate stats
points.foreach(binSeqOp(nodeStatsAggregators, _))
// transform nodeStatsAggregators array to (nodeIndex, nodeAggregateStats) pairs,
// which can be combined with other partition using `reduceByKey`
nodeStatsAggregators.view.zipWithIndex.map(_.swap).iterator
}
}
首先对每个partition构造一个DTStatsAggregator数组,长度是node的个数,注意这里实际使用的是数组,node怎样与自己的aggregator的对应?前面我们提到NodeIndexInfo的第一个成员是groupIndex,其值就是node的次序,和这里aggregator数组index其实是对应的,也就是说可以从NodeIndexInfo中取得groupIndex,然后作为数组index取得对应node的agg。DTStatsAggregator的入参是metadata和每个node的特征子集。然后将每个点统计到DTStatsAggregator中,其中调用了内部函数binSeqOp,
/**
* Performs a sequential aggregation over a partition.
*
* 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 baggedPoint Data point being aggregated.
* @return agg
*/
def binSeqOp(
agg: Array[DTStatsAggregator],
baggedPoint: BaggedPoint[TreePoint]): Array[DTStatsAggregator] = {
//对每个node
treeToNodeToIndexInfo.foreach { case (treeIndex, nodeIndexToInfo) =>
val nodeIndex = predictNodeIndex(topNodes(treeIndex), baggedPoint.datum.binnedFeatures,
bins, metadata.unorderedFeatures)
nodeBinSeqOp(treeIndex, nodeIndexToInfo.getOrElse(nodeIndex, null), agg, baggedPoint)
}
agg
}
首先调用函数predictNodeIndex计算nodeIndex,如果是首轮或者叶子节点,直接返回node.id;如果不是首轮,因为传入的是每棵树的root node,就从root node开始,逐渐往下判断该point应该是属于哪个node的,因为我们已经对node进行了分裂,这里其实实现了样本的划分。举个栗子,当前node如果是root的左孩子节点,而point预测节点应该属于右孩子,则调用nodeBinSepOp时就直接返回了,不会将这个point统计进去,用不大的时间换取样本集划分的空间,还是比较巧妙的。
/**
* Get the node index corresponding to this data point.
* This function mimics prediction, passing an example from the root node down to a leaf
* or unsplit node; 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
* group of nodes on one call to [[findBestSplits()]].
*/
private def predictNodeIndex(
node: Node,
binnedFeatures: Array[Int],
bins: Array[Array[Bin]],
unorderedFeatures: Set[Int]): Int = {
if (node.isLeaf || node.split.isEmpty) {
// Node is either leaf, or has not yet been split.
node.id
} else {
//判断point属于当前node的左孩子还是右孩子
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) {
//下面还有完整的左右孩子node,递归判断
// 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)
}
}
}
}
然后调用nodeBinSeqOp函数
/**
* Performs a sequential aggregation over a partition for a particular tree and node.
*
* For each feature, the aggregate sufficient statistics are updated for the relevant
* bins.
*
* @param treeIndex Index of the tree that we want to perform aggregation for.
* @param nodeInfo The node info for the tree node.
* @param agg Array storing aggregate calculation, with a set of sufficient statistics
* for each (node, feature, bin).
* @param baggedPoint Data point being aggregated.
*/
def nodeBinSeqOp(
treeIndex: Int,
nodeInfo: RandomForest.NodeIndexInfo,
agg: Array[DTStatsAggregator],
baggedPoint: BaggedPoint[TreePoint]): Unit = {
if (nodeInfo != null) {
//node的groupIndex,见前文
val aggNodeIndex = nodeInfo.nodeIndexInGroup
//node使用的特征子集
val featuresForNode = nodeInfo.featureSubset
//取样本在这棵树中出现的次数 0/1/k
val instanceWeight = baggedPoint.subsampleWeights(treeIndex)
if (metadata.unorderedFeatures.isEmpty) {
orderedBinSeqOp(agg(aggNodeIndex), baggedPoint.datum, instanceWeight, featuresForNode)
} else {
mixedBinSeqOp(agg(aggNodeIndex), baggedPoint.datum, splits,
metadata.unorderedFeatures, instanceWeight, featuresForNode)
}
}
}
函数的入参是treeIndex,该node的NodeIndexInfo结构,所有node的累加器数组,样本。本函数是针对单个node的操作,这里可以看到取node对应的aggregator就是通过NodeIndexInfo的第一个成员nodeIndexInGroup作为agg数组的index。
如果不包含无序特征,调用orderedBinSeqOp函数
/**
* 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 (feature, bin).
* @param treePoint Data point being aggregated.
* @param instanceWeight Weight (importance) of instance in dataset.
*/
private def orderedBinSeqOp(
agg: DTStatsAggregator, //node的agg
treePoint: TreePoint,
instanceWeight: Double,
featuresForNode: Option[Array[Int]]): Unit = {
val label = treePoint.label
// Iterate over features.
if (featuresForNode.nonEmpty) {
// Use subsampled features
var featureIndexIdx = 0
while (featureIndexIdx < featuresForNode.get.size) {
//连续特征:离散化后的index
//离散特征:featureValue.toInt
val binIndex = treePoint.binnedFeatures(featuresForNode.get.apply(featureIndexIdx))
agg.update(featureIndexIdx, binIndex, label, instanceWeight)
featureIndexIdx += 1
}
} else {
// Use all features
val numFeatures = agg.metadata.numFeatures
var featureIndex = 0
while (featureIndex < numFeatures) {
val binIndex = treePoint.binnedFeatures(featureIndex)
agg.update(featureIndex, binIndex, label, instanceWeight)
featureIndex += 1
}
}
}
函数中区分了是否使用了全部特征,区别仅在于如果使用了部分特征(特征抽样),需要先在featuresForNode中取得特征的实际index。
函数其实就是取出样本的使用特征,特征值,label和weight,更新到aggregator中,更新逻辑我们在前文已经说明过了。
包含了无序离散特征,则使用mixedBinSeqOp,只有无序离散特征处理方法不同于orderedBinSeqOp函数
// Unordered feature
val featureValue = treePoint.binnedFeatures(featureIndex)
//找到特征值对应的allStats中的范围
//特征起始位置从featureOffsets中取得,长度是bins的个数乘以分类个数,2*(2^(M-1)-1)*statsSize,
//每一个split将样本集分成2部分,allStats中左边部分连续存放,右半部分连续存放,而不是左右一起存放。
//因此,左边的起始位置直接可以从featureOffsets中获取,右边起始位置是(2^(M-1)-1)*statsSize
val (leftNodeFeatureOffset, rightNodeFeatureOffset) =
agg.getLeftRightFeatureOffsets(featureIndexIdx)
// Update the left or right bin for each split.
val numSplits = agg.metadata.numSplits(featureIndex)
var splitIndex = 0
while (splitIndex < numSplits) {
//split中的categories中包含左半边特征值组合,splitIndex相当于其离散化后的特征index
if (splits(featureIndex)(splitIndex).categories.contains(featureValue)) {
agg.featureUpdate(leftNodeFeatureOffset, splitIndex, treePoint.label,
instanceWeight)
} else {
agg.featureUpdate(rightNodeFeatureOffset, splitIndex, treePoint.label,
instanceWeight)
}
splitIndex += 1
}
6.4.1.3. 全局统计
partitionAggregates.reduceByKey((a, b) => a.merge(b))
1
就是将所有存在allStats中的分区统计结果逐个对应相加得到全局统计结果。
6.4.2. bestSplits
获得所有的统计后,就可以遍历所有的特征,计算impurity gain,确定最佳的split。
val nodeToBestSplits = partitionAggregates.reduceByKey((a, b) => a.merge(b))
.map { case (nodeIndex, aggStats) =>
val featuresForNode = nodeToFeaturesBc.value.map { nodeToFeatures =>
nodeToFeatures(nodeIndex)
}
// find best split for each node
val (split: Split, stats: InformationGainStats, predict: Predict) =
binsToBestSplit(aggStats, splits, featuresForNode, nodes(nodeIndex))
(nodeIndex, (split, stats, predict))
}.collectAsMap()
对每个node其中调用了binsToBestSplit函数,下面进行详细说明。
6.4.2.1 init
函数首先获取node在树的第几层,树结构如图
树的id如图所示,判断node在第几层只需要判断id的二进制表示的最高位的1在第几位即可,比如6的二进制表示是110,最高位的1是在第3位,则其在第3层。
然后获取当前node的预测值和impurity
// calculate predict and impurity if current node is top node
val level = Node.indexToLevel(node.id)
var predictWithImpurity: Option[(Predict, Double)] = if (level == 0) {
None
} else {
Some((node.predict, node.impurity))
}
6.4.2.2 连续特征
对于连续特征而言,当取其某个特征值为best split后,node的样本会被分成大于该特征值和小于等于该特征值两部分,需要分别统计两部分的class分布情况;另一方面,我们查找best,因此要遍历所有特征值的情况,一种巧妙的方法是,从左边开始逐次累积统计数据,需要从某个特征值作为split时,当前累计值就是左边小于等于的情况,用最右的值减去左边就是右边的情况。
例如上图中的情况,是某特征6个特征值分布情况,第一行是左累计,第二行是原始分布,当以v2作为split时,左边的分布就是c0:8,c1:5,右边是v6的分布减去v2,c0:19-8=11,c1:14-5=9。
if (binAggregates.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.
//如上所述,累计
val nodeFeatureOffset = binAggregates.getFeatureOffset(featureIndexIdx)
var splitIndex = 0
while (splitIndex < numSplits) {
binAggregates.mergeForFeature(nodeFeatureOffset, splitIndex + 1, splitIndex)
splitIndex += 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)
//获得node的impurity,level==0时,需要根据当前class的分布计算
predictWithImpurity = Some(predictWithImpurity.getOrElse(
calculatePredictImpurity(leftChildStats, rightChildStats)))
val gainStats = calculateGainForSplit(leftChildStats,
rightChildStats, binAggregates.metadata, predictWithImpurity.get._2)
(splitIdx, gainStats)
}.maxBy(_._2.gain)
(splits(featureIndex)(bestFeatureSplitIndex), bestFeatureGainStats)
计算split分裂node的impurity增益时,调用了calculateGainForSplit函数,其中分别计算了左右的增益,然后概率合并,并计算了左右的预测值,代码比较简单,这里不再赘述。
6.4.2.3. Unordered categorical feature
只有获取左右class的统计情况方法不一致,其他是一样的。
6.4.2.4. Ordered categorical feature
对于连续特征,特征值或者是binIndex是有序的,或者说其数值可以排序,因此如果某个特征值被当做split,分隔的就是左右两部分;对于无序离散特征,其被split分隔后特征值属于哪个bin是确定的;对于有序离散特征,其特征值代表一定次序关系,但是不具有绝对大小的含义,其处理方法可以近似按照连续特征的方法处理,但是spark这里处理了下,可能更优点。
spark首先会确定一个centroid,然后特征会按这个排序,这个相当于连续特征的binIndex。例如centroid如果取每个特征值中class1的个数,假设有特征值0,1,2,3,class1的个数分别为4,2,1,3,其中如果按照连续特征的处理方法,假设用1作为node的分裂点,计算impurity gain的时候分成0,1和2,3两部分统计。如果按照centroid的方法,其特征值排序次序应该是2,1,3,0,以1作为分裂点,会被分成2,1和3,0两部分。
// Ordered categorical feature
val nodeFeatureOffset = binAggregates.getFeatureOffset(featureIndexIdx)
val numBins = binAggregates.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 = Range(0, numBins).map { case featureValue =>
val categoryStats = binAggregates.getImpurityCalculator(nodeFeatureOffset, featureValue)
val centroid = if (categoryStats.count != 0) {
if (binAggregates.metadata.isMulticlass) {
// For categorical variables in multiclass classification,
// the bins are ordered by the impurity of their corresponding labels.
categoryStats.calculate()
} else if (binAggregates.metadata.isClassification) {
// For categorical variables in binary classification,
// the bins are ordered by the count of class 1.
categoryStats.stats(1)
} else {
// For categorical variables in regression,
// the bins are ordered by the prediction.
categoryStats.predict
}
} else {
Double.MaxValue
}
(featureValue, centroid)
}
logDebug("Centroids for categorical variable: " + centroidForCategories.mkString(","))
// bins sorted by centroids
val categoriesSortedByCentroid = centroidForCategories.toList.sortBy(_._2)
上面的代码为不同的情况设置不同的centroid的选取方法。如果是多分类,使用impurity;如果是二分类,使用class1的个数;如果是回归,使用预测值(实际是均值)。然后将特征值按centroid重排序。
下面的处理基本与连续特征类似,先按排序次序累计,然后计算左右的impurity,计算impurity gain。由于要返回split,之前离散特征的split返回的是空Array,这里构造了split,第四个参数中加入了实际的特征值,类比unordered的情况。
计算完完所有的特征的gain,就可以选取最大增益时的split,最后collectAsMap,key是nodeIndex,value是split, InfomationGainStats,predict的三元组。
6.4.3. node分裂
计算完节点的best split,就要根据这个split进行node的分裂,包括当前节点的一些属性完善,左右孩子节点的构造等。
// Iterate over all nodes in this group.
nodesForGroup.foreach { case (treeIndex, nodesForTree) =>
nodesForTree.foreach { node =>
val nodeIndex = node.id
val nodeInfo = treeToNodeToIndexInfo(treeIndex)(nodeIndex)
val aggNodeIndex = nodeInfo.nodeIndexInGroup
//从刚刚计算的best split中获取相关数据
val (split: Split, stats: InformationGainStats, predict: Predict) =
nodeToBestSplits(aggNodeIndex)
logDebug("best split = " + split)
// Extract info for this node. Create children if not leaf.
//截止条件
val isLeaf = (stats.gain <= 0) || (Node.indexToLevel(nodeIndex) == metadata.maxDepth)
assert(node.id == nodeIndex)
node.predict = predict
node.isLeaf = isLeaf
node.stats = Some(stats)
node.impurity = stats.impurity
logDebug("Node = " + node)
//如果不是叶子节点,需要构造左右孩子节点
if (!isLeaf) {
node.split = Some(split)
//叶子节点的depth,当前level+1
val childIsLeaf = (Node.indexToLevel(nodeIndex) + 1) == metadata.maxDepth
//左右孩子节点是否是叶子节点
val leftChildIsLeaf = childIsLeaf || (stats.leftImpurity == 0.0)
val rightChildIsLeaf = childIsLeaf || (stats.rightImpurity == 0.0)
//构造左右孩子节点
node.leftNode = Some(Node(Node.leftChildIndex(nodeIndex),
stats.leftPredict, stats.leftImpurity, leftChildIsLeaf))
node.rightNode = Some(Node(Node.rightChildIndex(nodeIndex),
stats.rightPredict, stats.rightImpurity, rightChildIsLeaf))
if (nodeIdCache.nonEmpty) {
val nodeIndexUpdater = NodeIndexUpdater(
split = split,
nodeIndex = nodeIndex)
nodeIdUpdaters(treeIndex).put(nodeIndex, nodeIndexUpdater)
}
//如果不是叶子节点,加入到nodeQueue待分裂队列中
// enqueue left child and right child if they are not leaves
if (!leftChildIsLeaf) {
nodeQueue.enqueue((treeIndex, node.leftNode.get))
}
if (!rightChildIsLeaf) {
nodeQueue.enqueue((treeIndex, node.rightNode.get))
}
logDebug("leftChildIndex = " + node.leftNode.get.id +
", impurity = " + stats.leftImpurity)
logDebug("rightChildIndex = " + node.rightNode.get.id +
", impurity = " + stats.rightImpurity)
}
}
}
这里将当前节点的左右孩子节点继续加入nodeQueue中,这里面放的是需要继续分裂的节点,至此本轮的findBestSplits就完成了。
// Choose node splits, and enqueue new nodes as needed.
timer.start("findBestSplits")
DecisionTree.findBestSplits(baggedInput,
metadata, topNodes, nodesForGroup,
treeToNodeToIndexInfo, splits, bins, nodeQueue,
timer, nodeIdCache = nodeIdCache)
timer.stop("findBestSplits")
6.5. 循环训练
上节我们说到最后待分裂的节点会加入到nodeQueue中,回到RandomForest.run函数中
while (nodeQueue.nonEmpty) {
// Collect some nodes to split, and choose features for each node (if subsampling).
// Each group of nodes may come from one or multiple trees, and at multiple levels.
val (nodesForGroup, treeToNodeToIndexInfo) =
RandomForest.selectNodesToSplit(nodeQueue, maxMemoryUsage, metadata, rng)
// Sanity check (should never occur):
assert(nodesForGroup.size > 0,
s"RandomForest selected empty nodesForGroup. Error for unknown reason.")
// Choose node splits, and enqueue new nodes as needed.
timer.start("findBestSplits")
DecisionTree.findBestSplits(baggedInput, metadata, topNodes, nodesForGroup,
treeToNodeToIndexInfo, splits, bins, nodeQueue, timer, nodeIdCache = nodeIdCache)
timer.stop("findBestSplits")
}
当有非叶子节点不断加入nodeQueue中,这里不断分裂出节点,直到所有节点触发截止条件。
7. 构造随机森林
在上面的训练过程可以看到,从根节点topNode中不断向下分裂一直到触发截止条件就构造了一棵树所有的node,因此构造整个森林也是非常简单
//构造
val trees = topNodes.map(topNode => new DecisionTreeModel(topNode, strategy.algo))
//返回rf模型
new RandomForestModel(strategy.algo, trees)
8. 随机森林模型
8.1. TreeEnsembleModel
随机森林RandomForestModel继承自树集合模型TreeEnsembleModel
class TreeEnsembleModel(
protected val algo: Algo,
protected val trees: Array[DecisionTreeModel],
protected val treeWeights: Array[Double],
protected val combiningStrategy: EnsembleCombiningStrategy)
algo:Regression/Classification
trees:树数组
treeWeights:每棵树的权重,在RF中每棵树的权重是相同的,在Adaboost可能是不同的
combiningStrategy:树合并时的策略,Sum/Average/Vote,分类的话应该是Vote,RF应该是Average,GBDT应该是Sum。
sumWeights:成员变量,不在参数表中,是treeWeights的sum
预测函数
/**
* Predicts for a single data point using the weighted sum of ensemble predictions.
*
* @param features array representing a single data point
* @return predicted category from the trained model
*/
private def predictBySumming(features: Vector): Double = {
val treePredictions = trees.map(_.predict(features))
blas.ddot(numTrees, treePredictions, 1, treeWeights, 1)
}
将每棵树的预测结果与各自的weight向量相乘
/**
* Classifies a single data point based on (weighted) majority votes.
*/
private def predictByVoting(features: Vector): Double = {
val votes = mutable.Map.empty[Int, Double]
trees.view.zip(treeWeights).foreach { case (tree, weight) =>
val prediction = tree.predict(features).toInt
votes(prediction) = votes.getOrElse(prediction, 0.0) + weight
}
votes.maxBy(_._2)._1
}
将每棵树的预测class为key,将树的weight累加到Map中作为value,最后取权重和最大对应的class
8.2. RandomForestModel
RandomForestModel @Since("1.2.0") (
@Since("1.2.0") override val algo: Algo,
@Since("1.2.0") override val trees: Array[DecisionTreeModel])
extends TreeEnsembleModel(algo, trees, Array.fill(trees.length)(1.0),
combiningStrategy = if (algo == Classification) Vote else Average)
对于随机森林,其weight都是1,树合并策略如果是分类就是Vote,回归是Average。
模型生成后,如果要应用到线上,需要将训练后的模型保存下来,自己写代码解析模型文件,进行预测,因此要了解模型的保存和加载。
8.2.1. 模型保存
分为两部分,第一部分是metadata,保存了一些配置,包括模型名,模型版本,模型的algo是classification/regression,合并策略,每棵树的权重。
implicit val format = DefaultFormats
val ensembleMetadata = Metadata(model.algo.toString,
model.trees(0).algo.toString,
model.combiningStrategy.toString,
model.treeWeights)
val metadata = compact(render(
("class" -> className) ~ ("version" -> thisFormatVersion) ~
("metadata" -> Extraction.decompose(ensembleMetadata))))
sc.parallelize(Seq(metadata), 1).saveAsTextFile(Loader.metadataPath(path))
第二部分是随机森林的每棵树的保存
// Create Parquet data.
val dataRDD = sc.parallelize(model.trees.zipWithIndex).flatMap { case (tree, treeId) =>
tree.topNode.subtreeIterator.toSeq.map(node => NodeData(treeId, node))
}.toDF()
dataRDD.write.parquet(Loader.dataPath(path))
其中首先调用node的subtreeIterator函数,返回所有node的Iterator,然后转成DataFrame结构,再写成parquet格式的文件。我们来看subtreeIterator函数
/** Returns an iterator that traverses (DFS, left to right) the subtree of this node. */
private[tree] def subtreeIterator: Iterator[Node] = {
Iterator.single(this) ++ leftNode.map(_.subtreeIterator).getOrElse(Iterator.empty) ++
rightNode.map(_.subtreeIterator).getOrElse(Iterator.empty)
}
其实就是用前序遍历的方式返回了树中的所有node的Iterrator。
我们再来看NodeData,看看每个node保存了什么数据
def apply(treeId: Int, n: Node): NodeData = {
NodeData(treeId, n.id, PredictData(n.predict), n.impurity,
n.isLeaf, n.split.map(SplitData.apply), n.leftNode.map(_.id),
n.rightNode.map(_.id), n.stats.map(_.gain))
}
保存了node的预测值,impurity,是否是否叶子节点,split,左右孩子节点的id,gain。其中split中包含了特征id,特征阈值,特征类型,离散特征数组(其实就是Split结构)。
8.2.2. 模型加载
metadata的加载就是解析json,主要是树的重建
val trees = TreeEnsembleModel.SaveLoadV1_0.loadTrees(sc,
path, metadata.treeAlgo)
new RandomForestModel(Algo.fromString(metadata.algo), trees)
其中调用了loadTrees函数
/**
* Load trees for an ensemble, and return them in order.
* @param path path to load the model from
* @param treeAlgo Algorithm for individual trees (which may differ from the ensemble's
* algorithm).
*/
def loadTrees(
sc: SparkContext,
path: String,
treeAlgo: String): Array[DecisionTreeModel] = {
val datapath = Loader.dataPath(path)
val sqlContext = SQLContext.getOrCreate(sc)
val nodes = sqlContext.read.parquet(datapath).map(NodeData.apply)
val trees = constructTrees(nodes)
trees.map(new DecisionTreeModel(_, Algo.fromString(treeAlgo)))
}
先是读取数据文件,读成NodeData格式,然后调用constructTrees重建树结构
def constructTrees(nodes: RDD[NodeData]): Array[Node] = {
val trees = nodes
.groupBy(_.treeId)
.mapValues(_.toArray)
.collect()
.map { case (treeId, data) =>
(treeId, constructTree(data))
}.sortBy(_._1)
val numTrees = trees.size
val treeIndices = trees.map(_._1).toSeq
assert(treeIndices == (0 until numTrees),
s"Tree indices must start from 0 and increment by 1, but we found $treeIndices.")
trees.map(_._2)
}
主要功能按树的id分组后,调用constructTree重建树
/**
* Given a list of nodes from a tree, construct the tree.
* @param data array of all node data in a tree.
*/
def constructTree(data: Array[NodeData]): Node = {
val dataMap: Map[Int, NodeData] = data.map(n => n.nodeId -> n).toMap
assert(dataMap.contains(1),
s"DecisionTree missing root node (id = 1).")
constructNode(1, dataMap, mutable.Map.empty)
}
/**
* Builds a node from the node data map and adds new nodes to the input nodes map.
*/
private def constructNode(
id: Int,
dataMap: Map[Int, NodeData],
nodes: mutable.Map[Int, Node]): Node = {
if (nodes.contains(id)) {
return nodes(id)
}
val data = dataMap(id)
val node =
if (data.isLeaf) {
Node(data.nodeId, data.predict.toPredict, data.impurity, data.isLeaf)
} else {
val leftNode = constructNode(data.leftNodeId.get, dataMap, nodes)
val rightNode = constructNode(data.rightNodeId.get, dataMap, nodes)
val stats = new InformationGainStats(data.infoGain.get, data.impurity, leftNode.impurity,
rightNode.impurity, leftNode.predict, rightNode.predict)
new Node(data.nodeId, data.predict.toPredict, data.impurity, data.isLeaf,
data.split.map(_.toSplit), Some(leftNode), Some(rightNode), Some(stats))
}
nodes += node.id -> node
node
}
其实就是递归的从NodeData中获取数据,重建node
从上面的分析可以看到,spark保存模型使用了parquet格式,对于我们在别的环境中使用是非常不方便的,训练完模型后,我们可以参照spark的做法,按照前序遍历的方法以json的格式保存node,在别的环境下复建树结构就可以了。
9. 坑
特征id,样本是libsvm格式的,特征id从1开始,但是设置离散特征数categoricalFeaturesInfo需要从0开始,相当于样本特征id-1
离散特征值,一旦在categoricalFeaturesInfo中指定了特征值的个数k,spark会认为这个特征是从0开始,连续到k-1。如果其中特征不连续,特征数应该设置成最大特征值+1
对于连续特征,spark使用等频离散化方法,又对样本进行了抽样,效果其实很难保证,不知道作者是否比较过这种方法与等间隔离散化效果孰优孰劣
maxBins的设置需要考虑连续特征离散化效果,连续特征离散化值的个数是maxBins-1,同时maxBins必须大于categoricalFeaturesInfo中最大离散特征值的个数
ordered feature,之前的理解是有误的,这里的order仅仅是说这种特征是可以经过某种方式排列后变成有序,排序标准根据分类/回归而不同,在上面的文章有具体介绍。在我们的实践中,有的离散特征,例如薪资,1代表0-1000元,2代表1000-2000元,3代表2000-3000元,特征值的大小本身就表征了实际意义,这种应该直接按连续特征处理(当然也可以对比下效果决定)。
10. 结语
我们基本上是逐行分析了spark随机森林的实现,展现了其实现过程中使用的技巧,希望对大家在理解随机森林和其实现方法有所帮助。
---------------------
原文:https://blog.csdn.net/snaillup/article/details/72820346
本文深入分析Spark中随机森林算法的实现细节,涵盖决策树、特征处理、样本划分、训练流程及模型保存加载等方面,揭示其高效训练策略。
6510

被折叠的 条评论
为什么被折叠?



