SparkMLlib-----GMM算法
Gaussian Mixture Model(GMM)是一个很流行的聚类算法。它与K-Means的很像,但是K-Means的计算结果是算出每个数据点所属的簇,而GMM是计算出这些数据点分配到各个类别的概率。与K-Means对比K-Means存在一些缺点,比如K-Means的聚类结果易受样本中的一些极值点影响。此外GMM的计算结果由于是得出一个概率,得出一个概率包含的信息量要比简单的一个结果多,对于49%和51%的发生的事件如果仅仅使用简单的50%作为阈值来分为两个类别是非常危险的。
Gaussian Mixture Model,顾名思义,它是假设数据服从高斯混合分布,或者说是从多个高斯分布中生成出来的。每个GMM由K个高斯分布组成,每个高斯分布称为一个"Component",这些Component线性加在一起就组成了GMM的概率密度函数:
使用GMM做聚类的方法,我们先使用R等工具采样数据绘出数据点分布的图观察是否符合高斯混合分布,或者直接假设我们的数据是符合高斯混合分布的,之后根据数据推算出GMM的概率分布,对应的每个高斯分布就是每个类别,因为我们已知(假设)了概率密度分布的形式,要去求出其中参数,所以是一个参数估计的过程,我们要推导出每个混合成分的参数(均值向量mu,协方差矩阵sigma,权重weight),高斯混合模型在训练时使用了极大似然估计法,最大化以下对数似然函数:
该式无法直接解析求解,因此采用了期望-最大化方法(Expectation-Maximization,EM)方法求解,具体步骤如下:
1.根据给定的K值,初始化K个多元高斯分布以及其权重;
2.根据贝叶斯定理,估计每个样本由每个成分生成的后验概率;(EM方法中的E步)
3.根据均值,协方差的定义以及2步求出的后验概率,更新均值向量、协方差矩阵和权重;(EM方法的M步)重复2~3步,直到似然函数增加值已小于收敛阈值,或达到最大迭代次数
接下来进行模型的训练与分析,我们采用了mllib包封装的GMM算法,具体代码如下
package com.xj.da.gmm import breeze.linalg.DenseVector import breeze.numerics.sqrt import org.apache.commons.math.stat.correlation.Covariance import org.apache.spark.mllib.clustering.{GaussianMixture, GaussianMixtureModel} import org.apache.spark.mllib.linalg import org.apache.spark.mllib.linalg.distributed.RowMatrix import org.apache.spark.mllib.linalg.{Matrices, Matrix, Vectors} import org.apache.spark.mllib.stat.distribution.MultivariateGaussian import org.apache.spark.rdd.RDD import org.apache.spark.{SparkConf, SparkContext} import scala.collection.mutable.ArrayBuffer /** * author : kongcong * number : 27 * date : 2017/7/19 */ object GMMWithMultivariate { def main(args: Array[String]): Unit = { val conf = new SparkConf() //.setMaster("local") .setAppName("GMMWithMultivariate") val sc = new SparkContext(conf) val rawData: RDD[String] = sc.textFile("hdfs://master:8020/home/kongc/data/query_result.csv") //val rawData: RDD[String] = sc.textFile("data/query_result.csv") println("count: " + rawData.count()) //println(rawData.count()) // col1, col2, status val data: RDD[linalg.Vector] = rawData.map { line => val raw: Array[String] = line.split(",") Vectors.dense(raw(0).toDouble, raw(1).toDouble, raw(4).toDouble) } // data.collect().take(10).foreach(println(_)) // col1, col2, status val trainData: RDD[linalg.Vector] = rawData.map { line => val raw: Array[String] = line.split(",") Vectors.dense(raw(0).toDouble, raw(1).toDouble) } // trainData.collect().take(10).foreach(println(_)) // 指定初始模型 // 0 val filter0: RDD[linalg.Vector] = data.filter(_.toArray(2) == 0) println(filter0.count()) //23195 // 1 val filter1: RDD[linalg.Vector] = data.filter(_.toArray(2) == 1) println(filter1.count()) //14602 val w1: Double = (filter0.count()/319377.toDouble) val w2: Double = (filter1.count()/319377.toDouble) println(s"w1 = $w1") // 均值 val m0x: Double = filter0.map(_.toArray(0)).mean() val m0y: Double = filter0.map(_.toArray(1)).mean() val m1x: Double = filter1.map(_.toArray(0)).mean() val m1y: Double = filter1.map(_.toArray(1)).mean() // 方差 val vx0: Double = filter0.map(_.toArray(0)).variance() val vy0: Double = filter0.map(_.toArray(1)).variance() val vx1: Double = filter1.map(_.toArray(0)).variance() val vy1: Double = filter1.map(_.toArray(1)).variance() // 均值向量 val mu1: linalg.Vector = Vectors.dense(Array(m0x, m0y)) val mu2: linalg.Vector = Vectors.dense(Array(m1x, m1y)) println(s"mu1 : $mu1") println(s"mu2 : $mu2") val array: RDD[Array[Double]] = rawData.map { line => val raw: Array[String] = line.split(",") Array(raw(0).toDouble, raw(1).toDouble, raw(4).toDouble) } val f0: RDD[Array[Double]] = array.filter(_(2) == 0) val f1: RDD[Array[Double]] = array.filter(_(2) == 1) println("f0.count:"+f0.count()) println("f1.count:"+f1.count()) // 0 x,y求协方差矩阵 val x0: RDD[Double] = f0.map(_(0)) val y0: RDD[Double] = f0.map(_(1)) //println(x0.collect().length == y0.collect().length) // 1 x,y求协方差矩阵 val x1: RDD[Double] = f1.map(_(0)) val y1: RDD[Double] = f1.map(_(1)) val ma0: Array[Array[Double]] = Array(x0.collect(),y0.collect()) val ma1: Array[Array[Double]] = Array(x1.collect(),y1.collect()) val r0: RDD[Array[Double]] = sc.parallelize(ma0) val r1: RDD[Array[Double]] = sc.parallelize(ma1) val rdd0: RDD[linalg.Vector] = r0.map(f => Vectors.dense(f)) val rdd1: RDD[linalg.Vector] = r1.map(f => Vectors.dense(f)) val RM0: RowMatrix = new RowMatrix(rdd0) val RM1: RowMatrix = new RowMatrix(rdd1) // 计算协方差矩阵 //println(RM0.computeCovariance().numCols) /*val i: Double = DenseVector(1.0, 2.0, 3.0, 4.0) dot DenseVector(1.0, 1.0, 1.0, 1.0) val c0yx: Double = i - m0x * m0y*/ val c0yx: Double = DenseVector(x0.collect()) dot DenseVector(y0.collect()) - m0x * m0y val c1yx: Double = DenseVector(x1.collect()) dot DenseVector(y1.collect()) - m1x * m1y //cov(Vectors.dense(x0.collect()),Vectors.dense(y0.collect())) val sigma1 = Matrices.dense(2, 2, Array(vx0, c0yx, c0yx, vy0)) val sigma2 = Matrices.dense(2, 2, Array(vx1, c1yx, c1yx, vy1)) val gmm1 = new MultivariateGaussian(mu1, sigma1) val gmm2 = new MultivariateGaussian(mu2, sigma2) val gaussians = Array(gmm1, gmm2) // 构建一个GaussianMixtureModel需要两个参数 一个是权重数组 一个是组成混合高斯分布的每个高斯分布 val initModel = new GaussianMixtureModel(Array(w1, w2), gaussians) for (i <- 0 until initModel.k) { println("weight=%f\nmu=%s\nsigma=\n%s\n" format (initModel.weights(i), initModel.gaussians(i).mu, initModel.gaussians(i).sigma)) } val gaussianMixture = new GaussianMixture() val mixtureModel = gaussianMixture .setInitialModel(initModel) .setK(2) .setConvergenceTol(0.0001) .run(trainData) val predict: RDD[Int] = mixtureModel.predict(trainData) rawData.zip(predict).saveAsTextFile("hdfs://master:8020/home/kongc/data/out/gmm/predict2") for (i <- 0 until mixtureModel.k) { println("weight=%f\nmu=%s\nsigma=\n%s\n" format (mixtureModel.weights(i), mixtureModel.gaussians(i).mu, mixtureModel.gaussians(i).sigma)) } } }
参考:http://blog.pluskid.org/?p=39
http://dblab.xmu.edu.cn/blog/1456/