二分K-均值聚类

K-均值算法算法可以实现收敛，但是存在一个问题是，K-均值算法会收敛到局部最优解而不是全局最优。

一种用于度量聚类效果的指标是SSE（sum of Squared Error，误差平方和）。SSE值越小表示数据点越接近它们的质心，聚类效果也就越好。应为误差取了平方，因此更加重视那些远离中心的点。一种肯定可以降低SSE值的方法是增加簇的个数，但这违背了聚类的目标。聚类的目标是在簇数目不变的情况下提高簇的质量。

对K-均值进行改进的方案之一是对生成的簇进行后处理，一种方法是将具有最大SSE值的簇划分两个簇。

二分K-均值算法

概算发首先将所有点作为一个簇，然后将该簇一分为二。之后选择其中一个簇进行划分，选择哪一个簇进行划分取决于其划分结果是否可以最大程度降低SSE的值。上述基于SSE的划分过程不断重复，直到得到用户指定的簇的数目为止。

二分K-均值算法的伪代码形式如下：

• 将所有点看成一个簇
• 当簇数目小于k时：
  • 对每一个簇
    • 计算总误差
    • 在给定的簇上面进行K-均值划分（k=2）
    • 计算将该簇一分为二之后的总误差
  • 选择使得总误差最小的那个簇的进行划分

代码：

import numpy as np
import matplotlib.pyplot as plt

def loadDataSet(fileName):
    dataList = []
    with open(fileName) as fr:
        for line in fr.readlines():
            currentLine = line.strip().split('\t')
            floatLine = list(map(np.float, currentLine))
            dataList.append(floatLine)
    return np.array(dataList)


def calcEcludDist(vecA, vecB):
    return np.sqrt(np.sum(np.power(vecA - vecB, 2)))


def randCent(dataSet, k):
    n = np.shape(dataSet)[1]    #n: 特征数量
    centroids = np.zeros((k, n))  #centroids的维度（类别数， 特征数）
    for j in range(n):
        minJ = np.min(dataSet[:, j])
        rangeJ = float(np.max(dataSet[:, j]) - minJ)
        mean = (minJ + np.max(dataSet[:, j])) / 2
        centroids[:, j] = minJ + rangeJ * np.random.rand(k, 1).T
    return centroids


def KMeans(dataSet, k, distMeans=calcEcludDist, createCent = randCent):
    m = np.shape(dataSet)[0]
    clusterAssment = np.array(np.zeros((m, 2)))
    centroids = createCent(dataSet, k)
    clusterChanged = True
    while clusterChanged:
        clusterChanged = False
        for i in range(m):
            minDist = np.inf
            minIndex = -1
            for j in range(k):
                distJI = distMeans(centroids[j, :], dataSet[i, :])
                if distJI < minDist:
                    minDist = distJI
                    minIndex = j
            if clusterAssment[i, 0] != minIndex:
                clusterChanged = True
            clusterAssment[i, :] = minIndex, minDist ** 2
        for cent in range(k):
            ptsInClust = dataSet[np.nonzero(clusterAssment[:, 0] == cent)[0]]
            centroids[cent, :] = np.mean(ptsInClust, axis=0)
    return centroids, clusterAssment


def biKMeans(dataSet, k, distMeans = calcEcludDist):
    """

    :param dataSet:
    :param k: the number of expected classes
    :param distMeans: function that describes how to calculate distance
    :return:
    """
    global bestNewCents, bestCentToSplit, bestClustAss
    m = np.shape(dataSet)[0]  # m: the number of data set
    clusterAssment = np.array(np.zeros((m, 2))) # （类别，距离质心的距离平方）
    centroid0 = np.mean(dataSet, axis=0) # 选择均值作为最开始的质心
    centroidList = [centroid0]  # 质心向量
    for j in range(m):
        clusterAssment[j, 1] = distMeans(centroid0, dataSet[j, :]) ** 2 # 计算每个样本距离质心的距离平方
    while len(centroidList) < k:    # 如果当前的质心小于期望的质心数量（分类数量）
        lowestSSE = np.inf
        for i in range(len(centroidList)):
            ptsInCurrentCluster = dataSet[np.nonzero(clusterAssment[:, 0] == i)[0], :]  # 第i类的数据放入集合ptsInCurrentCluster
            newCentroids, splitClustAss = KMeans(ptsInCurrentCluster, 2, calcEcludDist) # 将第i类再分2类
            sseSplit = np.sum(splitClustAss[:, 1])  # 已分类数据的SSE总和
            sseNotSplit = np.sum(clusterAssment[np.nonzero(clusterAssment[:, 0] != i)[0], 1])   # 未分类数据的SSE总和
            print("sse split and not split: ", sseSplit, sseNotSplit)
            if sseSplit + sseNotSplit < lowestSSE:  # 如果SSE小于最小SSE，则说明应该对第i类数据再分类
                bestCentToSplit = i
                bestNewCents = newCentroids
                bestClustAss = splitClustAss.copy()
                lowestSSE = sseSplit + sseNotSplit
        bestClustAss[np.nonzero(bestClustAss[:, 0] == 1)[0], 0] = len(centroidList) # 选中的那类再分两类，如果类别为1,则新的类别号为len(centroidList)
        bestClustAss[np.nonzero(bestClustAss[:, 0] == 0)[0], 0] = bestCentToSplit   # 如果类别为0,则新的类别号为bestCenToSplit
        print("the bestCentToSplit is: ", bestCentToSplit)
        print("the len of bestClustAss is: ", len(bestClustAss))
        print("new cluster assessment: \n", bestClustAss)
        centroidList[bestCentToSplit] = bestNewCents[0, :]
        centroidList.append(bestNewCents[1, :])
        clusterAssment[np.nonzero(clusterAssment[:, 0] == bestCentToSplit)[0], :] = bestClustAss
    return centroidList, clusterAssment


if __name__ == "__main__":
    dataList = loadDataSet('testSet2.txt')
    sseSplit, sseAssments = biKMeans(dataList, 3)
    fig = plt.figure()
    ax = fig.add_subplot(111)
    ax.scatter(dataList[np.nonzero(sseAssments[:, 0] == 0)[0], 0], dataList[np.nonzero(sseAssments[:, 0] == 0)[0], 1], 
               c="blue")
    ax.scatter(dataList[np.nonzero(sseAssments[:, 0] == 1)[0], 0], dataList[np.nonzero(sseAssments[:, 0] == 1)[0], 1],
               c="red")
    ax.scatter(dataList[np.nonzero(sseAssments[:, 0] == 2)[0], 0], dataList[np.nonzero(sseAssments[:, 0] == 2)[0], 1],
               c="green")
    plt.show()

划分结果：

 

CONTACT INFORMATION

E-Mail: birdguan@seu.edu.cn

QQ: 46611253