基本Kmeans算法
import numpy as np
import random
def calc_dist(vec1, vec2):
"""计算两个向量的欧氏距离"""
return np.sqrt(sum(np.power(vec1-vec2, 2)))
def rand_cent(dataSet, k):
"""取k个随机质心,质心的每个特征的值是在数据上下界中随机选取
不一定是数据中包含的值"""
n = shape(dataSet)[1]
centroids = np.mat(np.zeros((k, n)))
for j in range(n):
min_val = min(dataSet[:, j])
range_val = max(dataSet[:, j]) - min_val
centroids[j] = min_val + range_val * random.rand(k, 1)
return centroids
def Kmeans(dataSet, k, dist_eval = calc_dist, create_cent = rand_cent):
"""
创建随机的K个点作为起始质心
当任意一个点的簇分配结果发生改变时:
对数据中的每个数据点:
对每个质心:
计算质心与数据点之间的距离
将数据点分配到距其最近的簇
对每个簇,计算簇中所有点的均值并将其作为质心
"""
m = shape(dataSet)[0]
cluster_assign = np.mat(np.zeros((m, 2))) #样本属于哪个簇和距离
centroids = create_cent(dataSet, k) #随机创建质心
cluster_change = True #程序终止条件
while cluster_change:
cluster_change = False
#更新每个样本的所在簇
for i in range(m):
min_dist = np.inf
min_index = -1
for j in range(k):
dist = dist_eval(dataSet[i, :], centroids[j, :])
if dist < min_dist:
min_dist = dist
min_index = j
if min_index != cluster_assign[i]:
cluster_change = True
cluster_assign[i] = [min_index, min_dist]
if not cluster_change:
break
#更新每个簇
for cent in range(k):
cluster_sample = dataSet[np.nonzero(cluster_assign[:, 0].A == cent)[0]]
centroids[cent, :] = np.mean(cluster_sample, axis = 0)
return centroids, cluster_assign
二分KMeans算法
def bin_Kmeans(dataSet, k, dist_eval = calc_dist):
"""
Kmeans容易收敛到局部最小值,为克服,有二分Kmeans:
1. 将所有点看成一个簇
2. 当簇数目小于k时继续划分:
对于每一个簇:
计算总误差
在给定的簇上面进行2Means聚类
计算将该簇一分为2之后的误差 + 其他簇类的误差作为新的总误差
选择使得总误差最小的的分割
"""
m = shape(dataSet)[0]
cluster_assign = np.mat(np.zeros((m, 2))) #第一列保存所属簇id, 第二列保存距离
centroid0 = np.mean(dataSet, axis = 0).tolist()[0] #初始簇为全数据集的中心
cent_list = [centroid0]
for i in range(m):
cluster_assign[i, 1] = dist_eval(np.mat(centroid0), dataSet[i, :]) ** 2 #用欧式距离的平方,更重视那些远离中心的点
while len(cent_list) < k:
lowest_sse = np.inf # sum of squared error
#选择最优分割簇,使总误差最小
for i in range(len(cent_list)):
cluster_samples = dataSet[np.nonzero(cluster_assign[:, 0].A == i)[0], :] #这个簇的所有样本
centroids, clusters = Kmeans(cluster_samples, 2, dist_eval) #对这个簇的样本一分为2
sse_other_cluster = sum(cluster_assign[np.nonzero(cluster_assign[:, 0].A != i)[0], 1]) #其他簇类的误差
sse_split = sum(clusters[:, 1]) #划分部分的误差
if sse_split + sse_other_cluster < lowest_sse:
lowest_sse = sse_split + sse_other_cluster
best_new_cnets = centroids
best_clusters = clusters
best_split_cent = i
#更新簇的分配结果
best_clusters[np.nonzero(best_clusters[:, 0].A == 1)[0], 0] = len(cent_list) #新的簇id
best_clusters[np.nonzero(best_clusters[:, 0].A == 0)[0], 0] = best_split_cent #被分割的簇id
cent_list[best_split_cent] = best_new_cnets[0:] #更新簇的特征值
cent_list.append(best_new_cnets[1:])
cluster_assign[np.nonzero(cluster_assign[:, 0].A == best_split_cent)[0], :] = best_clusters #更新总的样本所属簇记录
return np.mat(cent_list), cluster_assign