聚类——WKFCM的matlab程序
作者:凯鲁嘎吉 - 博客园 http://www.cnblogs.com/kailugaji/
在聚类——WKFCM文章中已介绍了WKFCM算法的理论知识,现在用matlab进行实现,下面这个例子是用FCM初始化聚类中心,也可以随机初始化聚类中心。
1.matlab程序
WKFCM_main.m
%function [ave_acc_WKFCM,max_acc_WKFCM,min_acc_WKFCM,ave_iter_WKFCM,ave_run_time]=WKFCM_main(X,real_label,K) function [ave_acc_WKFCM,max_acc_WKFCM,min_acc_WKFCM,ave_iter_FCM,ave_iter_WKFCM,ave_run_time]=WKFCM_main(X,real_label,K) %输入K:聚的类,max_iter是最大迭代次数,T:遗传算法最大迭代次数,n:种群个数 X:未归一化 %输出ave_acc_KFCM:迭代max_iter次之后的平均准确度,iter:实际KFCM迭代次数 t0=cputime; max_iter=20; s=0; s_1=0; s_2=0; accuracy=zeros(max_iter,1); iter_WKFCM_t=zeros(max_iter,1); iter_FCM_t=zeros(max_iter,1); %对data做最大-最小归一化处理 % [data_num,~]=size(data); % X=(data-ones(data_num,1)*min(data))./(ones(data_num,1)*(max(data)-min(data))); for i=1:max_iter %[label,~,iter_WKFCM]=My_WKFCM(X,K); [label,~,iter_WKFCM,iter_FCM]=My_WKFCM(X,K); iter_WKFCM_t(i)=iter_WKFCM; iter_FCM_t(i)=iter_FCM; accuracy(i)=succeed(real_label,K,label); s=s+accuracy(i); s_1=s_1+ iter_WKFCM_t(i); s_2=s_2+ iter_FCM_t(i); %fprintf('第 %2d 次,WKFCM的迭代次数为:%2d,准确度为:%.8f ', i, iter_WKFCM_t(i), accuracy(i)); fprintf('第 %2d 次,FCM的迭代次数为:%2d,WKFCM的迭代次数为:%2d,准确度为:%.8f ', i, iter_FCM_t(i), iter_WKFCM_t(i), accuracy(i)); end ave_iter_FCM=s_2/max_iter; ave_iter_WKFCM=s_1/max_iter; ave_acc_WKFCM=s/max_iter; max_acc_WKFCM=max(accuracy); min_acc_WKFCM=min(accuracy); run_time=cputime-t0; ave_run_time=run_time/max_iter;
My_WKFCM.m
%function [label_1,para_miu,iter]=My_WKFCM(X,K) function [label_1,para_miu,iter,iter_FCM]=My_WKFCM(X,K) %输入K:聚类数 %输出:label_1:聚的类, para_miu_new:模糊聚类中心μ,responsivity:模糊隶属度 format long eps=1e-4; %定义迭代终止条件的eps %sigma_2=2^(-4); %高斯核函数的参数sigma^2 sigma_2=150; %高斯核函数的参数sigma^2 beta=2; alpha=2; %模糊加权指数,[1,+无穷) T=100; %最大迭代次数 fitness=zeros(T,1); [X_num,X_dim]=size(X); distant=zeros(X_num,K,X_dim); kernel_fun=zeros(X_num,K,X_dim); R_temp=zeros(X_num,K,X_dim); miu_up=zeros(X_num,K,X_dim); miu_down=zeros(X_num,K,X_dim); W_temp=zeros(X_num,K,X_dim); J_temp=zeros(X_num,K,X_dim); count=zeros(X_num,1); %统计distant中每一行为0的个数 responsivity=zeros(X_num,K); R_up=zeros(X_num,K); W_up=zeros(K,X_dim); %---------------------------------------------------------------------------------------------------- %随机初始化属性权重K*X_dim para_weight=ones(K,X_dim)./X_dim; %随机初始化K个聚类中心 % rand_array=randperm(X_num); %产生1~X_num之间整数的随机排列 % para_miu=X(rand_array(1:K),:); %随机排列取前K个数,在X矩阵中取这K行作为初始聚类中心 %用FCM初始聚类中心 [~,para_miu,iter_FCM]=My_FCM(X,K); % ---------------------------------------------------------------------------------------------------- % WKFCM算法 for t=1:T %计算隶属函数K*X_num for j=1:X_dim for i=1:X_num for k=1:K distant(i,k,j)=(X(i,j)-para_miu(k,j))^2; kernel_fun(i,k,j)=exp((-distant(i,k,j))/sigma_2); R_temp(i,k,j)=(para_weight(k,j)^beta)*(1-kernel_fun(i,k,j)); end end end R_down=sum(R_temp,3); for i=1:X_num count(i)=sum(R_down(i,:)==0); if count(i)>0 for k=1:K if R_down(i,k)==0 responsivity(i,k)=1./count(i); else responsivity(i,k)=0; end end else R_up(i,:)=R_down(i,:).^(-1/(alpha-1)); %隶属度矩阵的分子部分N*K responsivity(i,:)= R_up(i,:)./sum( R_up(i,:),2); end end %更新聚类中心K*X_dim for j=1:X_dim for i=1:X_num for k=1:K miu_up(i,k,j)=responsivity(i,k)*kernel_fun(i,k,j)*X(i,j); miu_down(i,k,j)=responsivity(i,k)*kernel_fun(i,k,j); end end end miu_up_sum=sum(miu_up,1); miu_down_sum=sum(miu_down,1); for k=1:K for j=1:X_dim if para_weight(k,j)==0 para_miu(k,j)=0; else para_miu(k,j)=miu_up_sum(1,k,j)/miu_down_sum(1,k,j); end end end %更新属性权重K*X_dim for j=1:X_dim for i=1:X_num for k=1:K distant(i,k,j)=(X(i,j)-para_miu(k,j))^2; kernel_fun(i,k,j)=exp((-distant(i,k,j))./sigma_2); W_temp(i,k,j)=(responsivity(i,k)^alpha)*(1-kernel_fun(i,k,j)); end end end W_down=sum(W_temp,1); for k=1:K for j=1:X_dim if W_down(1,k,j)==0 para_weight(k,j)=1./X_dim; else W_up(k,:)=W_down(1,k,:).^(-1/(beta-1)); %属性权重矩阵的分子部分K*X_dim para_weight(k,:)= W_up(k,:)./sum( W_up(k,:),2); end end end %计算目标函数值 for j=1:X_dim for i=1:X_num for k=1:K distant(i,k,j)=(X(i,j)-para_miu(k,j))^2; kernel_fun(i,k,j)=exp((-distant(i,k,j))./sigma_2); J_temp(i,k,j)=(responsivity(i,k)^alpha)*(para_weight(k,j)^beta)*(1-kernel_fun(i,k,j)); end end end fitness(t)=2*sum(sum(sum( J_temp))); if t>1 if abs(fitness(t)-fitness(t-1))<eps break; end end end iter=t; %实际迭代次数 [~,label_1]=max(responsivity,[],2);
2.在UCI数据库的iris上的运行结果
>> [ave_acc_WKFCM,max_acc_WKFCM,min_acc_WKFCM,ave_iter_FCM,ave_iter_WKFCM,ave_run_time]=WKFCM_main(data,real_label,3) 第 1 次,FCM的迭代次数为:14,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 2 次,FCM的迭代次数为:17,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 3 次,FCM的迭代次数为:28,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 4 次,FCM的迭代次数为:14,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 5 次,FCM的迭代次数为:20,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 6 次,FCM的迭代次数为:11,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 7 次,FCM的迭代次数为:19,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 8 次,FCM的迭代次数为:15,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 9 次,FCM的迭代次数为:14,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 10 次,FCM的迭代次数为:11,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 11 次,FCM的迭代次数为:21,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 12 次,FCM的迭代次数为:20,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 13 次,FCM的迭代次数为:10,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 14 次,FCM的迭代次数为:28,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 15 次,FCM的迭代次数为:18,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 16 次,FCM的迭代次数为:16,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 17 次,FCM的迭代次数为:12,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 18 次,FCM的迭代次数为:20,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 19 次,FCM的迭代次数为:12,WKFCM的迭代次数为: 4,准确度为:0.92666667 第 20 次,FCM的迭代次数为:13,WKFCM的迭代次数为: 4,准确度为:0.92666667 ave_acc_WKFCM = 0.926666666666666 max_acc_WKFCM = 0.926666666666667 min_acc_WKFCM = 0.926666666666667 ave_iter_FCM = 16.649999999999999 ave_iter_WKFCM = 4 ave_run_time = 0.232812500000000