static_SVM_line_mean_var.zip_SVM 参数_mean_核函数寻优_核参数 优_线性核SVM

function [output]=static_SVM_line_mean_var(stacon,nconn) %% estimating classified accuracy with SVM clc; nSubjs=size(stacon,2); Con_MDD=stacon;%6672*38 %a=zeros(1,nconn); %co_Index=zeros(nconn,nSubjs);%30个特征集的肯德尔等级相关系数的排序序号 lab=Con_MDD(6671,:); best=zeros(nSubjs,4); addpath('D:\Software\libsvm-3.12\matlab'); d=1; feature=10:10:1000; output=cell(length(feature),2); for num_feature=feature for r=1:nSubjs Con_MDD_test=Con_MDD(:,r); Con_MDD_train=Con_MDD(:,setdiff(1:nSubjs,r)); MDD=Con_MDD_train(:,find(Con_MDD_train(6671,:)==1)); Con=Con_MDD_train(:,find(Con_MDD_train(6671,:)==-1)); train_label=Con_MDD_train(6671,:)'; test_label=Con_MDD_test(6671,:); %% 根据均值和方差进行特征选择 [train_data,test_data]=feature_mean_var(Con_MDD_train,Con_MDD_test,Con,MDD,nconn,num_feature); %% 参数寻优 option = struct('s',0,'t',0,'cmin',-14,'cmax',0,'w1min',-2,'w1max',4, ...%w1表示病人，w2表示正常人 'w2min',-2,'w2max',4,'gmin',-14,'gmax',0,'v',37,'cstep',0.5,'gstep',0.5,'w1step',0.5,'w2step',0.5); %[bestacc,bestc,bestg,bestw1,bestw2]=wangge_SVMcgForClass(train_label,train_data,option); [bestacc,bestc,bestw1,bestw2,cg]=SVMcw_line(train_label,train_data,option); best(r,:)=[bestacc,bestc,bestw1,bestw2]; %% 进行预测 cmd_temp = ['-s ',num2str(option.s),' -t ',num2str(option.t),' -c ',num2str( bestc ),' -w1 ',num2str(bestw1),' -w-1 ',num2str(bestw2)]; model=svmtrain(train_label,train_data, cmd_temp); [pred_lab(r),accuracy,dec_valued]=svmpredict(test_label,test_data,model); disp(['num_feature=',num2str(num_feature),' 第',num2str(r),'次循环']); end label=[lab;pred_lab]; matr=zeros(2,2); for i=1:length(lab) matr(lab(i)/2+3/2,pred_lab(i)/2+3/2)=matr(lab(i)/2+3/2,pred_lab(i)/2+3/2)+1; end right=[matr(1,1),matr(2,2)]; accuracy=sum(diag(matr))/sum(sum(matr)); output{d,1}=[num_feature,accuracy,right]; output{d,2}=best; d=d+1; end rmpath('D:\Software\libsvm-3.12\matlab'); end function [train_data,test_data]=feature_mean_var(Con_MDD_train,Con_MDD_test,Con,MDD,nconn,num_feature) %% 通过基于均值和方差的比较来进行特征选择 Con_data=Con';%person*feature MDD_data=MDD'; n1=size(Con_data,1);%正常人的样本数 n2=size(MDD_data,1);%病人的样本数 %计算各个特征的方差和 S=(var(Con_data)/n1+var(MDD_data)/n2).^0.5; %计算各个特征的重要性 Con_mean=mean(Con_data,1); MDD_mean=mean(MDD_data,1); for i=1:nconn F(i)=abs(Con_mean(i)-MDD_mean(i))/S(i); end [f,Index]=sort(F,'descend'); train_data=Con_MDD_train(Index(1:num_feature),:)'; test_data=Con_MDD_test(Index(1:num_feature),:)'; end function [bestacc,bestc,bestw1,bestw2,cg]=SVMcw_line(train_label,train_data,option) %网格参数寻优函数(分类问题):SVMcgForClass %输入： %train_label:训练集的标签，格式要求与svmtrain相同。 %train:训练集，格式要求与svmtrain相同。 %cmin,cmax:惩罚参数c的变化范围，即在[2^cmin,2^cmax]范围内寻找最佳的参数c，默认值为cmin=-8，cmax=8，即默认惩罚参数c的范围是[2^(-8),2^8]。 %gmin,gmax:RBF核参数g的变化范围，即在[2^gmin,2^gmax]范围内寻找最佳的RBF核参数g，默认值为gmin=-8，gmax=8，即默认RBF核参数g的范围是[2^(-8),2^8]。 %v:进行Cross Validation过程中的参数，即对训练集进行v-fold Cross Validation，默认为3，即默认进行3折CV过程。 %cstep,gstep:进行参数寻优是c和g的步进大小，即c的取值为2^cmin,2^(cmin+cstep),…,2^cmax,，g的取值为2^gmin,2^(gmin+gstep),…,2^gmax，默认取值为cstep=1,gstep=1。 %accstep:最后参数选择结果图中准确率离散化显示的步进间隔大小（[0,100]之间的一个数），默认为4.5。 %输出： %bestCVaccuracy:最终CV意义下的最佳分类准确率。 %bestc:最佳的参数c。 %bestg:最佳的参数g。 %% X:c A:w1 B;w2 cg:CVaccuracy train=train_data; X=option.cmin:option.cstep:option.cmax; [A,B] =meshgrid(option.w1min:option.w1step:option.w1max,... option.w2min:option.w2step:option.w2max);%产生网格矩阵 [m,n] =size(A); %[X,Y] =meshgrid(cmin:cstep:cmax,gmin:gstep:gmax);%产生网格矩阵 %[m,n] =size(X); cg=cell(length(X),1); c_g =zeros(m,n); eps = 10^(-4); d=1; %% record acc with different c & w,and find the bestacc with the smallest c bestc = 1; %bestg = 0.1; bestacc = 0; basenum = 2; bestw1=1; bestw2=0.5; for t=X for i = 1:m for j = 1:n cmd = ['-s ',num2str(option.s),' -t ',num2str(option.t),' -v ',num2str(option.v),' -c ',num2str( basenum^t ),... ' -w1 ',num2str( basenum^A(i,j) ),' -w-1 ',num2str( basenum^B(i,j) )]; c_g(i,j) = svmtrain(train_label, train, cmd); if c_g(i,j) <= 65 continue; end if c_g(i,j) > bestacc bestacc = c_g(i,j); bestc = basenum^t; %bestg = basenum^Y(i,j); bestw1=basenum^A(i,j); bestw2=basenum^B(i,j); end if abs( c_g(i,j)-bestacc )<=eps && bestc > basenum^t bestacc = c_g(i,j); bestc = basenum^t; %bestg = basenum^Y(i,j); bestw1=basenum^A(i,j); bestw2=basenum^B(i,j); end end end cg{d}=c_g; d=d+1; end end