模糊算法 c语言,模糊C均值聚类算法

最近在学习粗糙集相关知识。模糊集作为相似的领域也一并了解。在应用方面，模糊集在图像分割方面应用比较广。中文资料中，模糊C均值聚类算法的资料最为丰富，因此使用模糊C均值聚类算法(FCM)作为学习的起点吧。

模糊C均值聚类算法的原理在这篇博客中讲的非常清楚。

算法原理详解.

原理的主要难点是对目标函数求偏导以获得极值点。因此算法的问题之一就是极值点不一定是最小值。

为了方便对比，我参照代码原地址重新封装了代码，并使用numpy对运行效率进行了优化。使之尽量和sklearn的调用方法相同。代码如下：

# -*- coding: utf-8 -*-

"""

Created on Wed Jul 3 09:26:30 2019

@author: BHN

"""

import numpy as np

class FCM():

def __init__(self,n_clusters):

super(FCM, self).__init__()

self.n_clusters = n_clusters

self.u = None

self.core = None

def initialize_U(self,data_num):

u = np.random.rand(data_num, self.n_clusters)

data_max = np.sum(u, axis=1)

# print(data_max)

data_max = np.tile(data_max, (self.n_clusters,1)).T

u = np.true_divide(u,data_max)

self.u = u

return u

def fit(self,data,m=2,max_item=100):

self.u = self.initialize_U(data.shape[0])

for i in range(max_item):

############################更新聚类中心############################

new_core = []

for data_dimension in range(self.n_clusters):

u_tmp = np.tile(self.u[:,data_dimension],(data.shape[1],1)).T

u_tmp = np.power(u_tmp,m)

new_core.append(np.true_divide(np.sum(np.multiply(u_tmp,data),axis=0),\

np.sum(u_tmp,axis=0)))

new_core = np.array(new_core)

self.core = new_core

# print(new_core)

#############################计算各点与聚类中心的距离################

new_core_tmp = new_core.reshape((new_core.shape[0],1,new_core.shape[1]))

new_core_tmp = np.tile(new_core_tmp,((1,data.shape[0],1)))

data_tmp = data.reshape((1,data.shape[0],data.shape[1]))

data_tmp = np.tile(data_tmp,((self.n_clusters,1,1)))

distance_tmp = np.subtract(new_core_tmp,data_tmp)

distance_tmp = np.power(distance_tmp,2)

distance_tmp = np.sum(distance_tmp,axis=2)

distance_tmp = np.power(distance_tmp,1/2).T

#############################更新隶属矩阵###########################

tmp1 = distance_tmp.reshape((distance_tmp.shape[0],distance_tmp.shape[1],1))

tmp1 = np.tile(tmp1,(self.n_clusters))

tmp2 = np.tile(distance_tmp,(self.n_clusters))

tmp2 = tmp2.reshape((distance_tmp.shape[0],distance_tmp.shape[1],distance_tmp.shape[1]))

new_u = np.true_divide(tmp1,tmp2)

new_u = np.power(new_u,1/(m-1))

new_u = np.sum(new_u,axis=2)

new_u = np.power(new_u,-1)

#############################查看迭代是否终止########################

subtract_2_u = np.sum(np.fabs(np.subtract(self.u,new_u)))

self.u = new_u

# print(subtract_2_u)

if subtract_2_u == 0:

break

# print(new_u)

def fit_predict(self,data,m=2,max_item=1000):

ans = []

if self.u == None:

self.fit(data,m=m,max_item=max_item)

ans = np.argmax(self.u, axis=1)

return ans

if __name__ =='__main__':

cls = FCM(n_clusters=2)

# cls.fit(np.random.rand(5, 3),2)

ans = cls.fit_predict(np.random.rand(5, 3),2)

print(ans)

最后和k-mean算法进行对比发现在我对比的几种数据分布中，FCM算法的聚类效果要差于K-mean。原因可能是因为隶属函数的应用并不能帮助其改善聚类情况吧。两种算法的效果对比图如下。

导出的图片.jpg