分享

Spark MLlib之 KMeans聚类算法详解

问题导读

1.什么是Spark MLlib ?
2.Spark MLlib 分为哪些类?
3.KMeans算法的基本思想是什么?
4.Spark Mllib KMeans源码包含哪些内容?



一直想学习下Spark 的机器学习,今天总结整理下。


1.什么是Spark MLlib

MLlib 是Spark对常用的机器学习算法的实现库,同时包括相关的测试和数据生成器。



2.Spark MLlib 分类
MLlib 目前支持四种常见的机器学习问题:二元分类,回归,聚类以及协同过滤,同时也包括一个底层的梯度下降优化基础算法。

我们知道了分类,这里重点介绍聚类


3.KMeans算法的基本思想

KMeans算法的基本思想是初始随机给定K个簇中心,按照最邻近原则把待分类样本点分到各个簇。然后按平均法重新计算各个簇的质心,从而确定新的簇心。一直迭代,直到簇心的移动距离小于某个给定的值。


K-Means聚类算法主要分为三个步骤:

(1)第一步是为待聚类的点寻找聚类中心;

(2)第二步是计算每个点到聚类中心的距离,将每个点聚类到离该点最近的聚类中去;

(3)第三步是计算每个聚类中所有点的坐标平均值,并将这个平均值作为新的聚类中心;

反复执行(2)、(3),直到聚类中心不再进行大范围移动或者聚类次数达到要求为止。



4.过程演示
下图展示了对n个样本点进行K-means聚类的效果,这里k取2:
(a)未聚类的初始点集;
(b)随机选取两个点作为聚类中心;
(c)计算每个点到聚类中心的距离,并聚类到离该点最近的聚类中去;
(d)计算每个聚类中所有点的坐标平均值,并将这个平均值作为新的聚类中心;
(e)重复(c),计算每个点到聚类中心的距离,并聚类到离该点最近的聚类中去;
(f)重复(d),计算每个聚类中所有点的坐标平均值,并将这个平均值作为新的聚类中心。

3.png


5.Spark Mllib KMeans源码分析

class KMeansprivate (
    privatevar k: Int,
    privatevar maxIterations: Int,
    privatevar runs: Int,
    privatevar initializationMode: String,
    privatevar initializationSteps: Int,
    privatevar epsilon: Double,
    privatevar seed: Long)extends Serializablewith Logging {
// KMeans类参数:
k:聚类个数,默认2;maxIterations:迭代次数,默认20;runs:并行度,默认1;
initializationMode:初始中心算法,默认"k-means||";initializationSteps:初始步长,默认5;epsilon:中心距离阈值,默认1e-4;seed:随机种子。
  /**
   * Constructs a KMeans instance with default parameters: {k: 2, maxIterations: 20, runs: 1,
   * initializationMode: "k-means||", initializationSteps: 5, epsilon: 1e-4, seed: random}.
   */
  defthis() =this(2,20, 1, KMeans.K_MEANS_PARALLEL,5, 1e-4, Utils.random.nextLong())
// 参数设置
/** Set the number of clusters to create (k). Default: 2. */
  def setK(k: Int):this.type = {
    this.k = k
    this
  }
**省略各个参数设置代码**
// run方法,KMeans主入口函数
  /**
   * Train a K-means model on the given set of points; `data` should be cached for high
   * performance, because this is an iterative algorithm.
   */
  def run(data: RDD[Vector]): KMeansModel = {

    if (data.getStorageLevel == StorageLevel.NONE) {
      logWarning("The input data is not directly cached, which may hurt performance if its"
        + " parent RDDs are also uncached.")
    }

// Compute squared norms and cache them.
// 计算每行数据的L2范数,数据转换:data[Vector]=> data[(Vector, norms)],其中norms是Vector的L2范数,norms就是:。
    val norms = data.map(Vectors.norm(_,2.0))
    norms.persist()
    val zippedData = data.zip(norms).map {case (v, norm) =>
      new VectorWithNorm(v, norm)
    }
    val model = runAlgorithm(zippedData)
    norms.unpersist()

    // Warn at the end of the run as well, for increased visibility.
    if (data.getStorageLevel == StorageLevel.NONE) {
      logWarning("The input data was not directly cached, which may hurt performance if its"
        + " parent RDDs are also uncached.")
    }
    model
  }
// runAlgorithm方法,KMeans实现方法。
  /**
   * Implementation of K-Means algorithm.
   */
  privatedef runAlgorithm(data: RDD[VectorWithNorm]): KMeansModel = {

    val sc = data.sparkContext

    val initStartTime = System.nanoTime()

    val centers =if (initializationMode == KMeans.RANDOM) {
      initRandom(data)
    } else {
      initKMeansParallel(data)
    }

    val initTimeInSeconds = (System.nanoTime() - initStartTime) /1e9
    logInfo(s"Initialization with $initializationMode took " +"%.3f".format(initTimeInSeconds) +
      " seconds.")

    val active = Array.fill(runs)(true)
    val costs = Array.fill(runs)(0.0)

    var activeRuns =new ArrayBuffer[Int] ++ (0 until runs)
    var iteration =0

    val iterationStartTime = System.nanoTime()
//KMeans迭代执行,计算每个样本属于哪个中心点,中心点累加样本的值及计数,然后根据中心点的所有的样本数据进行中心点的更新,并比较更新前的数值,判断是否完成。其中runs代表并行度。
    // Execute iterations of Lloyd's algorithm until all runs have converged
    while (iteration < maxIterations && !activeRuns.isEmpty) {
      type WeightedPoint = (Vector, Long)
      def mergeContribs(x: WeightedPoint, y: WeightedPoint): WeightedPoint = {
        axpy(1.0, x._1, y._1)
        (y._1, x._2 + y._2)
      }

      val activeCenters = activeRuns.map(r => centers(r)).toArray
      val costAccums = activeRuns.map(_ => sc.accumulator(0.0))

      val bcActiveCenters = sc.broadcast(activeCenters)

      // Find the sum and count of points mapping to each center
//计算属于每个中心点的样本,对每个中心点的样本进行累加和计算;
runs代表并行度,k中心点个数,sums代表中心点样本累加值,counts代表中心点样本计数;
contribs代表((并行度I,中心J),(中心J样本之和,中心J样本计数和));
findClosest方法:找到点与所有聚类中心最近的一个中心;
      val totalContribs = data.mapPartitions { points =>
        val thisActiveCenters = bcActiveCenters.value
        val runs = thisActiveCenters.length
        val k = thisActiveCenters(0).length
        val dims = thisActiveCenters(0)(0).vector.size

        val sums = Array.fill(runs, k)(Vectors.zeros(dims))
        val counts = Array.fill(runs, k)(0L)

        points.foreach { point =>
          (0 until runs).foreach { i =>
           val (bestCenter, cost) = KMeans.findClosest(thisActiveCenters(i), point)
           costAccums(i) += cost
           val sum = sums(i)(bestCenter)
           axpy(1.0, point.vector, sum)
           counts(i)(bestCenter) += 1
          }
        }

        val contribs =for (i <-0 until runs; j <-0 until k) yield {
          ((i, j), (sums(i)(j), counts(i)(j)))
        }
        contribs.iterator
      }.reduceByKey(mergeContribs).collectAsMap()
//更新中心点,更新中心点= sum/count;
判断newCenter与centers之间的距离是否 > epsilon * epsilon;
      // Update the cluster centers and costs for each active run
      for ((run, i) <- activeRuns.zipWithIndex) {
        var changed =false
        var j =0
        while (j < k) {
          val (sum, count) = totalContribs((i, j))
          if (count !=0) {
           scal(1.0 / count, sum)
           val newCenter =new VectorWithNorm(sum)
           if (KMeans.fastSquaredDistance(newCenter, centers(run)(j)) > epsilon * epsilon) {
             changed = true
           }
           centers(run)(j) = newCenter
          }
          j += 1
        }
        if (!changed) {
          active(run) = false
          logInfo("Run " + run +" finished in " + (iteration +1) + " iterations")
        }
        costs(run) = costAccums(i).value
      }

      activeRuns = activeRuns.filter(active(_))
      iteration += 1
    }

    val iterationTimeInSeconds = (System.nanoTime() - iterationStartTime) /1e9
    logInfo(s"Iterations took " +"%.3f".format(iterationTimeInSeconds) +" seconds.")

    if (iteration == maxIterations) {
      logInfo(s"KMeans reached the max number of iterations: $maxIterations.")
    } else {
      logInfo(s"KMeans converged in $iteration iterations.")
    }

    val (minCost, bestRun) = costs.zipWithIndex.min

    logInfo(s"The cost for the best run is $minCost.")

    new KMeansModel(centers(bestRun).map(_.vector))
  }
//findClosest方法:找到点与所有聚类中心最近的一个中心;
/**
   * Returns the index of the closest center to the given point, as well as the squared distance.
   */
  private[mllib]def findClosest(
      centers: TraversableOnce[VectorWithNorm],
      point: VectorWithNorm): (Int, Double) = {
    var bestDistance = Double.PositiveInfinity
    var bestIndex =0
    var i =0
    centers.foreach { center =>
      // Since `\|a - b\| \geq |\|a\| - \|b\||`, we can use this lower bound to avoid unnecessary
      // distance computation.
      var lowerBoundOfSqDist = center.norm - point.norm
      lowerBoundOfSqDist = lowerBoundOfSqDist * lowerBoundOfSqDist
      if (lowerBoundOfSqDist < bestDistance) {
        val distance: Double = fastSquaredDistance(center, point)
        if (distance < bestDistance) {
          bestDistance = distance
          bestIndex = i
        }
      }
      i += 1
    }
    (bestIndex, bestDistance)
  }
findClosest方法中:var lowerBoundOfSqDist = center.norm - point.norm
lowerBoundOfSqDist = lowerBoundOfSqDist * lowerBoundOfSqDist
如果中心点center是(a1,b1),需要计算的点point是(a2,b2),那么lowerBoundOfSqDist是:
4.png
如下是展开式,第二个是真正计算欧式距离时的除去开平方的公式。(在查找最短距离的时候无需计算开方,因为只需要计算出开方里面的式子就可以进行比较了,mllib也是这样做的)
5.png
可轻易证明上面两式的第一式将会小于等于第二式,因此在进行距离比较的时候,先计算很容易计算的lowerBoundOfSqDist,如果lowerBoundOfSqDist都不小于之前计算得到的最小距离bestDistance,那真正的欧式距离也不可能小于bestDistance了,因此这种情况下就不需要去计算欧式距离,省去很多计算工作。
如果lowerBoundOfSqDist小于了bestDistance,则进行距离的计算,调用fastSquaredDistance,这个方法将调用MLUtils.scala里面的fastSquaredDistance方法,计算真正的欧式距离,代码如下:
/**
   * Returns the squared Euclidean distance between two vectors. The following formula will be used
   * if it does not introduce too much numerical error:
   * <pre>
   *   \|a - b\|_2^2 = \|a\|_2^2 + \|b\|_2^2 - 2 a^T b.
   * </pre>
   * When both vector norms are given, this is faster than computing the squared distance directly,
   * especially when one of the vectors is a sparse vector.
   *
   * @param v1 the first vector
   * @param norm1 the norm of the first vector, non-negative
   * @param v2 the second vector
   * @param norm2 the norm of the second vector, non-negative
   * @param precision desired relative precision for the squared distance
   * @return squared distance between v1 and v2 within the specified precision
   */
  private[mllib]def fastSquaredDistance(
      v1: Vector,
      norm1: Double,
      v2: Vector,
      norm2: Double,
      precision: Double = 1e-6): Double = {
    val n = v1.size
    require(v2.size == n)
    require(norm1 >= 0.0 && norm2 >=0.0)
    val sumSquaredNorm = norm1 * norm1 + norm2 * norm2
    val normDiff = norm1 - norm2
    var sqDist =0.0
    /*
     * The relative error is
     * <pre>
     * EPSILON * ( \|a\|_2^2 + \|b\\_2^2 + 2 |a^T b|) / ( \|a - b\|_2^2 ),
     * </pre>
     * which is bounded by
     * <pre>
     * 2.0 * EPSILON * ( \|a\|_2^2 + \|b\|_2^2 ) / ( (\|a\|_2 - \|b\|_2)^2 ).
     * </pre>
     * The bound doesn't need the inner product, so we can use it as a sufficient condition to
     * check quickly whether the inner product approach is accurate.
     */
    val precisionBound1 =2.0 * EPSILON * sumSquaredNorm / (normDiff * normDiff + EPSILON)
    if (precisionBound1 < precision) {
      sqDist = sumSquaredNorm - 2.0 * dot(v1, v2)
    } elseif (v1.isInstanceOf[SparseVector] || v2.isInstanceOf[SparseVector]) {
      val dotValue = dot(v1, v2)
      sqDist = math.max(sumSquaredNorm - 2.0 * dotValue,0.0)
      val precisionBound2 = EPSILON * (sumSquaredNorm +2.0 * math.abs(dotValue)) /
        (sqDist + EPSILON)
      if (precisionBound2 > precision) {
        sqDist = Vectors.sqdist(v1, v2)
      }
    } else {
      sqDist = Vectors.sqdist(v1, v2)
    }
    sqDist
  }
fastSquaredDistance方法会先计算一个精度,有关精度的计算val precisionBound1 = 2.0 * EPSILON * sumSquaredNorm / (normDiff * normDiff + EPSILON),如果在精度满足条件的情况下,欧式距离sqDist = sumSquaredNorm - 2.0 * v1.dot(v2),sumSquaredNorm即为 6.png ,2.0 * v1.dot(v2)即为 7.png 。这也是之前将norm计算出来的好处。如果精度不满足要求,则进行原始的距离计算公式了 8.png ,即调用Vectors.sqdist(v1, v2)。


6.Mllib KMeans实例

1、数据
数据格式为:特征1 特征2 特征3
0.0 0.0 0.0
0.1 0.1 0.1
0.2 0.2 0.2
9.0 9.0 9.0
9.1 9.1 9.1
9.2 9.2 9.2

2、代码
  //1读取样本数据
  valdata_path ="/home/jb-huangmeiling/kmeans_data.txt"
  valdata =sc.textFile(data_path)
  valexamples =data.map { line =>
    Vectors.dense(line.split(' ').map(_.toDouble))
  }.cache()
  valnumExamples =examples.count()
  println(s"numExamples = $numExamples.")
  //2建立模型
  valk =2
  valmaxIterations =20
  valruns =2
  valinitializationMode ="k-means||"
  valmodel = KMeans.train(examples,k, maxIterations,runs, initializationMode)
  //3计算测试误差
  valcost =model.computeCost(examples)
  println(s"Total cost = $cost.")



参考:
Spark MLlib KMeans聚类算法
作者:sunbow0


已有(2)人评论

跳转到指定楼层
pzz 发表于 2017-3-27 11:34:38
请问有随机森林算法的程序吗,能否发一份给我。
回复

使用道具 举报

符伯云 发表于 2020-10-29 11:15:11
麻烦问一下您的spark是哪个版本的
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则

关闭

推荐上一条 /2 下一条