Github user mengxr commented on a diff in the pull request:
https://github.com/apache/spark/pull/216#discussion_r10947573
--- Diff:
mllib/src/main/scala/org/apache/spark/mllib/discretization/EntropyMinimizationDiscretizer.scala
---
@@ -0,0 +1,402 @@
+/*
+ * 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.discretization
+
+import scala.collection.mutable.Stack
+import org.apache.spark.SparkContext._
+import org.apache.spark.mllib.util.InfoTheory
+import org.apache.spark.rdd.RDD
+import org.apache.spark.storage.StorageLevel
+import org.apache.spark.mllib.regression.LabeledPoint
+import scala.collection.mutable
+
+
+/**
+ * This class contains methods to discretize continuous values with the
method proposed in
+ * [Fayyad and Irani, Multi-Interval Discretization of Continuous-Valued
Attributes, 1993]
+ */
+class EntropyMinimizationDiscretizer private (
+ @transient var data: RDD[LabeledPoint],
+ @transient var continousFeatures: Seq[Int],
+ var elementsPerPartition: Int = 18000,
+ var maxBins: Int = Int.MaxValue)
+ extends DiscretizerModel {
+
+ private var thresholds = Map.empty[Int, Seq[Double]]
+ private val partitions = { x : Long => math.ceil(x.toDouble /
elementsPerPartition).toInt }
+
+ def this() = this(null, null)
+
+ /**
+ * Sets the RDD[LabeledPoint] to be discretized
+ *
+ * @param data RDD[LabeledPoint] to be discretized
+ */
+ def setData(data: RDD[LabeledPoint]): EntropyMinimizationDiscretizer = {
+ this.data = data
+ this
+ }
+
+ /**
+ * Sets the indexes of the features to be discretized
+ *
+ * @param continuousFeatures Indexes of features to be discretized
+ */
+ def setContinuousFeatures(continuousFeatures: Seq[Int]):
EntropyMinimizationDiscretizer = {
+ this.continousFeatures = continuousFeatures
+ this
+ }
+
+ /**
+ * Sets the maximum number of elements that a partition should have
+ *
+ * @param ratio Maximum number of elements for a partition
+ * @return The Discretizer with the parameter changed
+ */
+ def setElementsPerPartition(ratio: Int): EntropyMinimizationDiscretizer
= {
+ this.elementsPerPartition = ratio
+ this
+ }
+
+ /**
+ * Sets the maximum number of discrete values
+ *
+ * @param maxBins Maximum number of discrete values
+ * @return The Discretizer with the parameter changed
+ */
+ def setMaxBins(maxBins: Int): EntropyMinimizationDiscretizer = {
+ this.maxBins = maxBins
+ this
+ }
+
+ /**
+ * Returns the thresholds used to discretized the given feature
+ *
+ * @param feature The number of the feature to discretize
+ */
+ def getThresholdsForFeature(feature: Int): Seq[Double] = {
+ thresholds.get(feature) match {
+ case Some(a) => a
+ case None =>
+ val featureValues = data.map({
+ case LabeledPoint(label, values) => (values(feature),
label.toString.trim)
+ })
+ val sortedValues = featureValues.sortByKey()
+ val initial_candidates = initialThresholds(sortedValues)
+ val thresholdsForFeature = this.getThresholds(initial_candidates)
+ this.thresholds += ((feature, thresholdsForFeature))
+ thresholdsForFeature
+ }
+ }
+
+ /**
+ * Returns the thresholds used to discretized the given features
+ *
+ * @param features The number of the feature to discretize
+ */
+ def getThresholdsForFeature(features: Seq[Int]): Map[Int, Seq[Double]] =
{
+ for (feature <- features diff this.thresholds.keys.toSeq) {
+ getThresholdsForFeature(feature)
+ }
+
+ this.thresholds.filter({ features.contains(_) })
+ }
+
+ /**
+ * Returns the thresholds used to discretized the continuous features
+ */
+ def getThresholdsForContinuousFeatures: Map[Int, Seq[Double]] = {
+ for (feature <- continousFeatures diff this.thresholds.keys.toSeq) {
+ getThresholdsForFeature(feature)
+ }
+
+ this.thresholds
+ }
+
+ /**
+ * Calculates the initial candidate treholds for a feature
+ * @param data RDD of (value, label) pairs
+ * @return RDD of (candidate, class frequencies between last and current
candidate) pairs
+ */
+ private def initialThresholds(data: RDD[(Double, String)]): RDD[(Double,
Map[String,Int])] = {
+ data.mapPartitions({ it =>
+ var lastX = Double.NegativeInfinity
+ var lastY = ""
+ var result = Seq.empty[(Double, Map[String, Int])]
+ var freqs = Map.empty[String, Int]
+
+ for ((x, y) <- it) {
+ if (x != lastX && y != lastY && lastX != Double.NegativeInfinity) {
+ // new candidate and interval
+ result = ((x + lastX) / 2, freqs) +: result
+ freqs = freqs.empty + ((y, 1))
+ } else {
+ // we continue on the same interval
+ freqs = freqs.updated(y, freqs.getOrElse(y, 0) + 1)
+ }
+ lastX = x
+ lastY = y
+ }
+
+ // we add last element as a candidate threshold for convenience
+ result = (lastX, freqs) +: result
+
+ result.reverse.toIterator
+ }).persist(StorageLevel.MEMORY_AND_DISK)
+ }
+
+ /**
+ * Returns a sequence of doubles that define the intervals to make the
discretization.
+ *
+ * @param candidates RDD of (value, label) pairs
+ */
+ private def getThresholds(candidates: RDD[(Double, Map[String,Int])]):
Seq[Double] = {
+
+ //Create queue
+ val stack = new mutable.Queue[((Double, Double), Option[Double])]
+
+ //Insert first in the stack
+ stack.enqueue(((Double.NegativeInfinity, Double.PositiveInfinity),
None))
+ var result = Seq(Double.NegativeInfinity)
+
+ // While more elements to eval, continue
+ while(stack.length > 0 && result.size < this.maxBins){
+
+ val (bounds, lastThresh) = stack.dequeue
+
+ var cands = candidates.filter({ case (th, _) => th > bounds._1 && th
<= bounds._2 })
+ val nCands = cands.count
+ if (nCands > 0) {
+ cands = cands.coalesce(partitions(nCands))
+
+ evalThresholds(cands, lastThresh) match {
+ case Some(th) =>
+ result = th +: result
+ stack.enqueue(((bounds._1, th), Some(th)))
+ stack.enqueue(((th, bounds._2), Some(th)))
+ case None => /* criteria not fulfilled, finish */
+ }
+ }
+ }
+ (Double.PositiveInfinity +: result).sorted
+ }
+
+ /**
+ * Selects one final thresholds among the candidates and returns two
partitions to recurse
+ *
+ * @param candidates RDD of (candidate, class frequencies between last
and current candidate)
+ * @param lastSelected last selected threshold to avoid selecting it
again
+ */
+ private def evalThresholds(
+ candidates: RDD[(Double, Map[String, Int])],
+ lastSelected : Option[Double]) = {
+
+ var result = candidates.map({
+ case (cand, freqs) =>
+ (cand, freqs, Seq.empty[Int], Seq.empty[Int])
+ }).cache
+
+ val numPartitions = candidates.partitions.size
+ val bc_numPartitions = candidates.context.broadcast(numPartitions)
+
+ // stores accumulated freqs from left to right
+ val l_total = candidates.context.accumulator(Map.empty[String,
Int])(MapAccumulator)
+ // stores accumulated freqs from right to left
+ val r_total = candidates.context.accumulator(Map.empty[String,
Int])(MapAccumulator)
+
+ // calculates accumulated frequencies for each candidate
+ (0 until numPartitions) foreach { l2r_i =>
+
+ val bc_l_total = l_total.value
+ val bc_r_total = r_total.value
+
+ result =
+ result.mapPartitionsWithIndex({ (slice, it) =>
+
+ val l2r = slice == l2r_i
+ val r2l = slice == bc_numPartitions.value - 1 - l2r_i
+
+ if (l2r && r2l) {
+
+ // accumulate both from left to right and right to left
+ var partialResult = Seq.empty[(Double, Map[String, Int],
Seq[Int], Seq[Int])]
+ var accum = Map.empty[String, Int]
+
+ for ((cand, freqs, _, r_freqs) <- it) {
+ accum = Utils.sumFreqMaps(accum, freqs)
+ val l_freqs = Utils.sumFreqMaps(accum,
bc_l_total).values.toList
+ partialResult = (cand, freqs, l_freqs, r_freqs) +:
partialResult
+ }
+
+ l_total += accum
+
+ val r2lIt = partialResult.iterator
+ partialResult = Seq.empty[(Double, Map[String, Int], Seq[Int],
Seq[Int])]
+ accum = Map.empty[String, Int]
+ for ((cand, freqs, l_freqs, _) <- r2lIt) {
+ val r_freqs = Utils.sumFreqMaps(accum,
bc_r_total).values.toList
+ accum = Utils.sumFreqMaps(accum, freqs)
+ partialResult = (cand, freqs, l_freqs, r_freqs) +:
partialResult
+ }
+ r_total += accum
+
+ partialResult.iterator
+
+ } else if (l2r) {
+
+ // accumulate freqs from left to right
+ var partialResult = Seq.empty[(Double, Map[String, Int],
Seq[Int], Seq[Int])]
+ var accum = Map.empty[String, Int]
+
+ for ((cand, freqs, _, r_freqs) <- it) {
+ accum = Utils.sumFreqMaps(accum, freqs)
+ val l_freqs = Utils.sumFreqMaps(accum,
bc_l_total).values.toList
+ partialResult = (cand, freqs, l_freqs, r_freqs) +:
partialResult
+ }
+
+ l_total += accum
+ partialResult.reverseIterator
+
+ } else if (r2l) {
+
+ // accumulate freqs from right to left
+ var partialResult = Seq.empty[(Double, Map[String, Int],
Seq[Int], Seq[Int])]
+ var accum = Map.empty[String, Int]
+ val r2lIt = it.toSeq.reverseIterator
+
+ for ((cand, freqs, l_freqs, _) <- r2lIt) {
+ val r_freqs = Utils.sumFreqMaps(accum,
bc_r_total).values.toList
+ accum = Utils.sumFreqMaps(accum, freqs)
+ partialResult = (cand, freqs, l_freqs, r_freqs) +:
partialResult
+ }
+
+ r_total += accum
+
+ partialResult.iterator
+
+ } else {
+ // do nothing in this iteration
+ it
+ }
+ }, true) // important to maintain partitions within the loop
+ .persist(StorageLevel.MEMORY_AND_DISK) // needed, otherwise spark
repeats calculations
+
+ result.foreachPartition({ _ => }) // Forces the iteration to be
calculated
+ }
+
+ // calculate h(S)
+ // s: number of elements
+ // k: number of distinct classes
+ // hs: entropy
+
+ val s = l_total.value.values.reduce(_ + _)
+ val hs = InfoTheory.entropy(l_total.value.values.toSeq, s)
+ val k = l_total.value.values.size
+
+ // select best threshold according to the criteria
+ val final_candidates =
+ result.flatMap({
+ case (cand, _, l_freqs, r_freqs) =>
+
+ val k1 = l_freqs.size
+ val s1 = if (k1 > 0) l_freqs.reduce(_ + _) else 0
+ val hs1 = InfoTheory.entropy(l_freqs, s1)
+
+ val k2 = r_freqs.size
+ val s2 = if (k2 > 0) r_freqs.reduce(_ + _) else 0
+ val hs2 = InfoTheory.entropy(r_freqs, s2)
+
+ val weighted_hs = (s1 * hs1 + s2 * hs2) / s
+ val gain = hs - weighted_hs
+ val delta = Utils.log2(3 ^ k - 2) - (k * hs - k1 * hs1 -
k2 * hs2)
+ var criterion = (gain - (Utils.log2(s - 1) + delta) / s) >
-1e-5
+
+ lastSelected match {
+ case None =>
+ case Some(last) => criterion = criterion && (cand != last)
+ }
+
+ if (criterion) {
+ Seq((weighted_hs, cand))
+ } else {
+ Seq.empty[(Double, Double)]
+ }
+ })
+
+ // choose best candidates and partition data to make recursive calls
+ if (final_candidates.count > 0) {
+ val selected_threshold = final_candidates.reduce({ case ((whs1,
cand1), (whs2, cand2)) =>
+ if (whs1 < whs2) (whs1, cand1) else (whs2, cand2)
+ })._2;
+ Some(selected_threshold)
+ } else {
+ None
+ }
+
+ }
+
+ /**
+ * Discretizes a value with a set of intervals.
+ *
+ * @param value The value to be discretized
+ * @param thresholds Thresholds used to asign a discrete value
+ */
+ private def assignDiscreteValue(value: Double, thresholds: Seq[Double])
= {
+ var aux = thresholds.zipWithIndex
+ while (value > aux.head._1) aux = aux.tail
+ aux.head._2
+ }
+
+ /**
+ * Discretizes an RDD of (label, array of values) pairs.
+ */
+ def discretize: RDD[LabeledPoint] = {
+ // calculate thresholds that aren't already calculated
+ getThresholdsForContinuousFeatures
+
+ val bc_thresholds = this.data.context.broadcast(this.thresholds)
+
+ // applies thresholds to discretize every continuous feature
+ data.map {
+ case LabeledPoint(label, values) =>
+ LabeledPoint(label,
+ values.zipWithIndex map {
+ case (value, i) =>
+ if (bc_thresholds.value.keySet contains i) {
+ assignDiscreteValue(value, bc_thresholds.value(i))
+ } else {
+ value
+ }
+ })
+ }
+ }
+
--- End diff --
Remove empty line.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
