聚类是机器学习中的无监督学习方法的重要一种，近来看了周志华老师的机器学习，专门研究了有关于聚类的一章，收获很多，对于其中的算法也动手实现了一下。主要实现的包括比较常见的k均值聚类、密度聚类和层次聚类，这三种聚类方法上原理都不难，算法过程也很清晰明白。有关于原理可以参阅周志华老师的机器学习第九章，这里只做一下代码的实现。
运行环境是Python2.7+numpy，说实话，numpy坑还是挺多的，其实用Matlab可能会更简单。
k均值聚类，核心是是不断更新簇样本的质心。
#encoding=utf-8
__author__ = 'freedom'
from numpy import*
import matplotlib.pyplot as plt
def loadDataSet(fileName):
'''
本函数用于加载数据
:param fileName: 数据文件名
:return:数据集，具有矩阵形式
'''
fr = open(fileName)
dataSet = []
for line in fr.readlines():
curLine = line.strip().split('\t')
inLine = map(float,curLine) # 利用map广播，是的读入的字符串变为浮点型
dataSet.append(inLine)
return mat(dataSet)
def getDistance(vecA,vecB):
'''
本函数用于计算欧氏距离
:param vecA: 向量A
:param vecB: 向量B
:return:欧氏距离
'''
return sqrt(sum(power(vecA-vecB,2)))
def randCent(dataSet,k):
'''
本函数用于生成k个随机质心
:param dataSet: 数据集，具有矩阵形式
:param k:指定的质心个数
:return:随机质心，具有矩阵形式
'''
n = shape(dataSet)[1] # 获取特征数目
centRoids = mat(zeros((k,n)))
for j in range(n):
minJ = min(dataSet[:,j]) # 获取每个特征的最小值
rangeJ = float(max(dataSet[:,j]-minJ)) # 获取每个特征的范围
centRoids[:,j] = minJ + rangeJ*random.rand(k,1) # numpy下的rand表示随机生成k*1的随机数矩阵，范围0-1
return centRoids
def kMeans(dataSet,k,disMens = getDistance,createCent = randCent):
'''
本函数用于k均值聚类
:param dataSet: 数据集，要求有矩阵形式
:param k: 指定聚类的个数
:param disMens: 求解距离的方式，除欧式距离还可以定义其他距离计算方式
:param createCent: 生成随机质心方式
:return:随机质心，簇索引和误差距离矩阵
'''
m = shape(dataSet)[0]
clusterAssment = mat(zeros((m,2))) # 要为每个样本建立一个簇索引和相对的误差，所以需要m行的矩阵，m就是样本数
centRoids = createCent(dataSet,k) # 生成随机质心
clusterChanged = True
while clusterChanged:
clusterChanged = False
for i in range(m): # 遍历所有样本
minDist = inf;minIndex = -1 # 初始化最小值
for j in range(k): # 遍历所有质心
disJI = disMens(centRoids[j,:],dataSet[i,:])
if disJI < minDist:
minDist = disJI;minIndex = j # 找出距离当前样本最近的那个质心
if clusterAssment[i,0] != minIndex: # 更新当前样本点所属于的质心
clusterChanged = True # 如果当前样本点不属于当前与之距离最小的质心，则说明簇分配结果仍需要改变
clusterAssment[i,:] = minIndex,minDist**2
for cent in range(k):
ptsInClust = dataSet[nonzero(clusterAssment[:,0].A == cent)[0]]
# nonzero 返回的是矩阵中所有非零元素的坐标，坐标的行数与列数个存在一个数组或矩阵当中
# 矩阵支持检查元素的操作，所有可以写成matrix == int这种形式，返回的一个布尔型矩阵，代表矩阵相应位置有无此元素
# 这里指寻找当前质心下所聚类的样本
centRoids[cent,:] = mean(ptsInClust,axis = 0) # 更新当前的质心为所有样本的平均值，axis = 0代表对列求平均值
return centRoids,clusterAssment
def plotKmens(dataSet,k,clusterMeans):
'''
本函数用于绘制kMeans的二维聚类图
:param dataSet: 数据集
:param k: 聚类的个数
:return:无
'''
centPoids,assment = clusterMeans(dataSet,k)
fig = plt.figure()
ax = fig.add_subplot(111)
ax.scatter(dataSet[:,0],dataSet[:,1],c = 'blue')
ax.scatter(centRoids[:,0],centRoids[:,1],c = 'red',marker = '+',s = 70)
plt.show()
def binKMeans(dataSet, k, distMeas = getDistance):
'''
本函数用于二分k均值算法
:param dataSet: 数据集，要求有矩阵形式
:param k: 指定聚类个数
:param distMeas: 求解距离的方式
:return:质心，簇索引和误差距离矩阵
'''
m = shape(dataSet)[0]
clusterAssment = mat(zeros((m,2)))
centRoids0 = mean(dataSet,axis = 0).tolist()[0] # 初始化一个簇，只有一个质心，分量就是就是所有特征的均值
# 注意，tolist函数用于将矩阵转化为一个列表，此列表为嵌套列表
#print centRoids0
centList = [centRoids0]
for j in range(m): # 遍历所有样本，计算所有样本与当前质心的距离作为误差
clusterAssment[j,1] = distMeas(mat(centRoids0),dataSet[j,:])**2
while (len(centList) < k): # 循环条件为当前质心数目还不够指定数目
lowestSSE = inf
for i in range(len(centList)): # 遍历所有质心
ptsCurrCluster = dataSet[nonzero(clusterAssment[:,0].A == i)[0],:] # 搜索到当前质心所聚类的样本
centroidsMat,splitClusterAss = kMeans(ptsCurrCluster,2,distMeas) # 将当前分割成两个簇
sseSplit = sum(splitClusterAss[:,1]) # 计算分裂簇后的SSE
sseNotSplit = sum(clusterAssment[nonzero(clusterAssment[:,0].A != i)[0],1])
# 计算分裂之前的SSE
if (sseSplit + sseNotSplit) < lowestSSE: # 如果分裂之后的SSE小，则更新
bestCent2Split = i
bestNewCents = centroidsMat
bestClustAss = splitClusterAss.copy()
lowestSSE = sseSplit+sseNotSplit
#重新编制簇的编号，凡是分裂后编号为1的簇，编号为质心列表长度，编号为0的簇，编号为最佳分裂质心的编号，以此更新
bestClustAss[nonzero(bestClustAss[:,0].A == 1)[0],0] = len(centList)
bestClustAss[nonzero(bestClustAss[:,0].A == 0)[0],0] = bestCent2Split
centList[bestCent2Split] = bestNewCents[0,:].tolist()[0] # 添加分裂的质心到质心列表中
centList.append(bestNewCents[1,:].tolist()[0])
clusterAssment[nonzero(clusterAssment[:,0].A == bestCent2Split)[0],:] = bestClustAss
return mat(centList),clusterAssment
def biKmeans(dataSet, k, distMeas=getDistance):
m = shape(dataSet)[0]
clusterAssment = mat(zeros((m,2)))
centroid0 = mean(dataSet, axis=0).tolist()[0]
centList =[centroid0] #create a list with one centroid
for j in range(m):#calc initial Error
clusterAssment[j,1] = distMeas(mat(centroid0), dataSet[j,:])**2
while (len(centList) < k):
lowestSSE = inf
for i in range(len(centList)):
ptsInCurrCluster = dataSet[nonzero(clusterAssment[:,0].A==i)[0],:]#get the data points currently in cluster i
centroidMat, splitClustAss = kMeans(ptsInCurrCluster, 2, distMeas)
sseSplit = sum(splitClustAss[:,1])#compare the SSE to the currrent minimum
sseNotSplit = sum(clusterAssment[nonzero(clusterAssment[:,0].A!=i)[0],1])
print "sseSplit, and notSplit: ",sseSplit,sseNotSplit
if (sseSplit + sseNotSplit) < lowestSSE:
bestCentToSplit = i
bestNewCents = centroidMat
bestClustAss = splitClustAss.copy()
lowestSSE = sseSplit + sseNotSplit
bestClustAss[nonzero(bestClustAss[:,0].A == 1)[0],0] = len(centList) #change 1 to 3,4, or whatever
bestClustAss[nonzero(bestClustAss[:,0].A == 0)[0],0] = bestCentToSplit
print 'the bestCentToSplit is: ',bestCentToSplit
print 'the len of bestClustAss is: ', len(bestClustAss)
centList[bestCentToSplit] = bestNewCents[0,:].tolist()[0]#replace a centroid with two best centroids
centList.append(bestNewCents[1,:].tolist()[0])
clusterAssment[nonzero(clusterAssment[:,0].A == bestCentToSplit)[0],:]= bestClustAss#reassign new clusters, and SSE
return mat(centList), clusterAssment
密度聚类，基本思路就是将所有密度可达的点都归为一簇。
#encoding=utf-8
import numpy as np
import kmeans as km
import matplotlib.pyplot as plt
def createDisMat(dataMat):
m = dataMat.shape[0]
n = dataMat.shape[1]
distMat = np.mat(np.zeros((m,m))) # 初始化距离矩阵，这里默认使用欧式距离
for i in range(m):
for j in range(m):
if i == j:
distMat[i,j] = 0
else:
dist = km.getDistance(dataMat[i,:],dataMat[j,:])
distMat[i,j] = dist
distMat[j,i] = dist
return distMat
def findCore(dataMat,delta,minPts):
core = []
m = dataMat.shape[0]
n = dataMat.shape[1]
distMat = createDisMat(dataMat)
for i in range(m):
temp = distMat[i,:] < delta # 单独抽取矩阵一行做过滤，凡是小于邻域值的都被标记位True类型
ptsNum = np.sum(temp,1) # 按行加和，统计小于邻域值的点个数
if ptsNum >= minPts:
core.append(i) # 满足条件，增加核心点
return core
def DBSCAN(dataMat,delta,minPts):
k = 0
m = dataMat.shape[0]
distMat = createDisMat(dataMat) # 获取距离矩阵
core = findCore(dataMat,delta,minPts) # 获取核心点列表
unVisit = [1] * m # hash值作为标记，当某一位置的数据位1时，表示还未被访问，为0表示已经被访问
Q = []
ck = []
unVistitOld = []
while len(core) != 0:
print 'a'
unVistitOld = unVisit[:] # 保留原始的未被访问集
i = np.random.choice(core) # 在核心点集中随机选择样本
Q.append(i) # 加入对列Q
unVisit[i] = 0 #剔除当前加入对列的数据，表示已经访问到了
while len(Q) != 0:
print len(Q)
temp = distMat[Q[0],:]<delta # 获取在此核心点邻域范围内的点集
del Q[0]
ptsNum = np.sum(temp,1)
if ptsNum >= minPts:
for j in range(len(unVisit)):
if unVisit[j] == 1 and temp[0,j] == True:
Q.append(j)
unVisit[j] = 0
k += 1
ck.append([])
for index in range(m):
if unVistitOld[index] == 1 and unVisit[index] == 0: # 上一轮未被访问到此轮被访问到的点均要加入当前簇
ck[k-1].append(index)
if index in core: # 在核心点集中清除当前簇的点
del core[core.index(index)]
return ck
def plotAns(dataSet,ck):
fig = plt.figure()
ax = fig.add_subplot(111)
ax.scatter(dataSet[ck[0],0],dataSet[ck[0],1],c = 'blue')
ax.scatter(dataSet[ck[1],0],dataSet[ck[1],1],c = 'red')
ax.scatter(dataSet[ck[2],0],dataSet[ck[2],1],c = 'green')
ax.scatter(dataSet[ck[3],0],dataSet[ck[3],1],c = 'yellow')
#ax.scatter(centRoids[:,0],centRoids[:,1],c = 'red',marker = '+',s = 70)
plt.show()
if __name__ == '__main__':
dataMat = km.loadDataSet("testSet.txt")
# distMat = createDisMat(dataMat)
# core = findCore(dataMat,1,5)
# print distMat
# print len(core)
ck = DBSCAN(dataMat,2,15)
print ck
print len(ck)
plotAns(dataMat,ck)
层次聚类，核心是定义了簇之间的距离衡量，不断寻找距离最近的簇归为一簇。
#encoding=utf-8
import numpy as np
import DBSCAN as db
import kmeans as km
def calcDistByMin(dataMat,ck1,ck2): # 最小距离点作为簇间的距离
min = np.inf
for vec1 in ck1:
for vec2 in ck2:
dist = km.getDistance(dataMat[vec1,:],dataMat[vec2,:])
if dist <= min:
min = dist
return min
def calcDistByMax(dataMat,ck1,ck2): # 最大距离点作为簇间的距离
max = 0
for vec1 in ck1:
for vec2 in ck2:
dist = km.getDistance(dataMat[vec1,:],dataMat[vec2,:])
if dist >= max:
max = dist
return max
def createDistMat(dataMat,calcDistType = calcDistByMin): # 生成初始的距离矩阵
m = dataMat.shape[0]
distMat = np.mat(np.zeros((m,m)))
for i in range(m):
for j in range(m):
listI = [i];listJ = [j] # 为配合距离函数的输入参数形式，在这里要列表化一下
distMat[i,j] = calcDistType(dataMat,listI,listJ)
distMat[j,i] = distMat[i,j]
return distMat
def findMaxLoc(distMat,q): # 寻找矩阵中最小的元素并返回其位置，注意，这里不能返回相同的坐标
min = np.inf
I = J = 0
for i in range(q):
for j in range(q):
if distMat[i,j] < min and i != j:
min = distMat[i,j]
I = i
J = j
return I,J
def ANGES(dataMat,k,calcDistType = calcDistByMax):
m = dataMat.shape[0]
ck = []
for i in range(m):
ck.append([i])
distMat = createDistMat(dataMat,calcDistType)
q = m # 初始化点集个数
while q > k:
i,j = findMaxLoc(distMat,q)
#print i,j
if i > j:
i,j = j,i # 保证i<j，这样做是为了删除的是序号较大的簇
ck[i].extend(ck[j]) # 把序号较大的簇并入序号小的簇
del ck[j] # 删除序号大的簇
distMat = np.delete(distMat,j,0) # 在距离矩阵中删除该簇的数据，注意这里delete函数有返回值，否则不会有删除作用
distMat = np.delete(distMat,j,1)
print distMat.shape
for index in range(0,q-1): # 重新计算新簇和其余簇之间的距离
distMat[i,index] = calcDistType(dataMat,ck[i],ck[index])
distMat[i,index] = distMat[index,i]
q -= 1 # 一个点被分入簇中，自减
return ck
if __name__ == '__main__':
dataMat = km.loadDataSet("testSet.txt")
ck = ANGES(dataMat,4)
print ck
db.plotAns(dataMat,ck)
