FCM公式详细推及代码

FCM作为一种模糊聚类的方法，依靠的是概率来进行聚类的，它的准则函数是：
$$\begin{gathered} J={\Sigma }_{j=1}^C{\Sigma }_{i=1}^N[{\mu }_j(x_i){]}^b||x_i-m_j|{|}^2 \\ st:{\Sigma }_{j=1}^C{\mu }_j(x_i)=1 \end{gathered}$$
之后需要将这个约束条件使用拉格朗日乘子放入到$J$中，得到的损失函数就是：
$$J={\Sigma }_{j=1}^C{\Sigma }_{i=1}^N[{\mu }_j(x_i){]}^b||x_i-m_j|{|}^2+{\Sigma }_{i=1}^N{\lambda }_i({\Sigma }_{j=1}^C{\mu }_j(x_i)-1)$$

介绍了这个之后，我们基于$J$分别对${\mu }_j(x_i),m_j$求偏导等于0：
$$\begin{gathered} \frac{\partial J}{\partial m_j}={\Sigma }_{i=1}^N{\mu }_j(x_i{)}^b(-2(x_i-m_j))=0 \\ 解得m_j=\frac{{\Sigma }_{i=1}^C{\mu }_j(x_i{)}^b{x}_i}{{\Sigma }_{i=1}^C{\mu }_j(x_i{)}^b} \end{gathered}$$
下面是对$\mu_j(x_i)求偏导数：
$$\begin{gathered} \frac{\partial J}{\partial {\mu }_j(x_i)}=b({\mu }_j(x_i{)}^{b-1}||x_i-m_j|{|}^2)+{\lambda }_i=0 \\ 解得{\mu }_j(x_i)=(\frac{-{\lambda }_i}{b}{)}^{\frac{1}{b-1}}||x_i-m_j|{|}^{-\frac{2}{(b-1)}} \end{gathered}$$
由于${\lambda }_i$不知道，但是知道${\Sigma }_{j=1}^C{\mu }_j(x_i)=1$,所以仍可以求解得到：
$${\mu }_j(x_i)=\frac{(\frac{-{\lambda }_i}{b}{)}^{\frac{1}{b-1}}||x_i-m_j|{|}^{-\frac{2}{(b-1)}}}{{\Sigma }_{k=1}^C(\frac{-{\lambda }_i}{b}{)}^{\frac{1}{b-1}}||x_i-m_k|{|}^{-\frac{2}{(b-1)}}}=\frac{||x_i-m_j|{|}^{-\frac{2}{(b-1)}}}{{\Sigma }_{k=1}^C||x_i-m_k|{|}^{-\frac{2}{(b-1)}}}$$
有了上面的这些准备工作之后，介绍FCM的整个工作流程：

1. 初始化参数$b,m_j$
2. 使用$m_j$来更新${\mu }_j(x_i)$
3. 使用新的${\mu }_j(x_i)$来更新$m_j$
4. 判断新的$m_j$和旧的$m_j$是否近似相等，若不相等返回step2;否则step5
5. 将样本根据${\mu }_j(x_i)$划分到隶属度最大的一类中

# -*- coding: utf-8 -*-
"""
FCM

@author: ASUS
"""
from sklearn import datasets
from sklearn.decomposition import PCA
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

def FCM(K,X,B):
    dist = np.array([[np.dot(X[i]- K[j],X[i]- K[j])**(-1/(B-1)) for j in range(len(K))] for i in range(len(X))])
    dist = np.sum(dist,axis = 1)
    pro = np.array([[np.dot(X[i]-K[j],X[i]-K[j])**(-1/(B-1))/dist[i] for j in range(len(K))]for i in range(len(X))])
    return pro.reshape(len(X),-1)

def precision(Y,Y_predict):
    return len(np.where(Y == Y_predict)[0])/len(Y)


iris_datas = datasets.load_iris()
Y = iris_datas.target
X_re = iris_datas.data
X = (X_re -np.min(X_re,axis = 0))/ (np.max(X_re,axis = 0)-np.min(X_re,axis = 0))     #归一化
#选取初始中心
K = np.array([np.average(X[50*i:50*(i+1)],axis = 0) for i in range(len(np.unique(Y)))])  #选取初始中心
Y_predict = np.zeros((len(Y),),dtype = np.uint32)                    #产生预测矩阵

#开始FCM的程序
B = 2
probility = np.ones((len(Y),len(K))) * 0.5
theta = 1
count = 0
while theta > 0.00001:
    probility = FCM(K,X,B)
    new_K = np.array([np.sum((probility[:,i]**(B)).reshape(len(X),-1)*X,axis = 0)/np.sum((probility[:,i])**B) for i in range(len(K))])
    theta = np.sum((K-new_K)**2)
    K = new_K
    count += 1

Y_predict = np.array([np.argmax(probility[i]) for i in range(len(X))])
print('FCM在iris数据集上的正确率为：'+str(round(precision(Y,Y_predict)*100,2))+'%')



#sonar数据集
sonar_datas = pd.read_csv('d:/microsoft/sonar.csv',header= None)
sonar_datas[61] = 0

sonar_datas.loc[np.where(sonar_datas[60]=='M')[0],61] = 1
    
Y = np.array(sonar_datas[61])
X = np.array(sonar_datas.iloc[:,:60])
X = (X -np.min(X,axis = 0))/ (np.max(X,axis = 0)-np.min(X,axis = 0))


# pca = PCA(0.95)
# X = pca.fit_transform(X)
K = np.array([np.average(X[:97],axis =0),np.average(X[97:],axis =0)])  #选取初始中心
#选取初始中心，用FCM中得到的聚类中心
Y_predict = np.zeros((len(Y),),dtype = np.uint32)
B = 2
probility = np.ones((len(Y),len(K))) * 0.5

theta = 1
count = 0
while theta > 0.00001:
    probility = FCM(K,X,B)
    new_K = np.array([np.sum((probility[:,i]**(B)).reshape(len(X),-1)*X,axis = 0)/np.sum((probility[:,i])**B) for i in range(len(K))])
    theta = np.sum((K-new_K)**2)
    K = new_K
    count += 1

Y_predict = np.array([np.argmax(probility[i]) for i in range(len(X))])
print('FCM在sonar数据集上的正确率为：'+str(round(precision(Y,Y_predict)*100,2))+'%')