基于pca的svm多分类器构造使用特征空间随机分割

clear all; clc; authors = 40 % number of authors txt = 30; traintxt = 0.9*txt; % number of each author's training set testtxt = txt-traintxt; classifier_num=100; validation_folds = 10 % cross validation's folds folds=validation_folds; fid3 = fopen(['E:\2011_research\ENexperiment\result\',num2str(authors),'_',num2str(txt),'_emsembles_pca+svm.txt'], 'w+'); % [label,allsamples]=libsvmread(['E:\2011_research\ENexperiment\ngrams\ngrams_',num2str(authors),'_',num2str(txt),'_','normalized_ig.libsvm']); % [label,allsamples]=libsvmread('E:\2011_research\ENexperiment\combine_20_30\all_20_10_normalized_ig.libsvm'); allsamples=zeros(authors*txt,20000); [label,allsamples]=libsvmread(['E:\2011_research\ENexperiment\combine_all\combine_',num2str(authors),'_',num2str(txt),'\all_normalized_ig_',num2str(authors),'_',num2str(txt),'.libsvm']);%30_30->888:85.67% initial_dimension = size(allsamples,2); tic load reduced_sample_10; now_dimension=size(reduced_sample,2) % save reduced_sample_20 reduced_sample; realauthor=0; count=0; accuracys=[]; subspace_size=now_dimension/classifier_num; for j=1:classifier_num space=0.2*now_dimension+randperm(0.8*now_dimension); % space=randperm(now_dimension); subspace(j,1:0.2*now_dimension)=1:0.2*now_dimension; subspace(j,0.2*now_dimension+1:0.6*now_dimension)=space(0.3*now_dimension+1:0.7*now_dimension); % subspace(j,:)=space(0.2*now_dimension+1:0.8*now_dimension); end % for j=1:classifier_num % [Sw, Sb, Sm]=scatter_mat(reduced_sample(:,subspace(j,:)),label,authors); % J3(j)=J3_comp(Sw, Sm); % end % [x,ind]=sort(J3); % maxJ3idx=rot90(ind); % n-folds cross validation base_correctnum=zeros(folds,classifier_num); indices = crossvalind('Kfold',label,10); for fold=1:folds testidx = (indices == fold); trainidx = ~testidx; traindata = reduced_sample(trainidx,:); traingroup = label(trainidx,:); trainlength = length(traingroup); testdata = reduced_sample(testidx,:); testgroup = label(testidx,:); testlength = length(testgroup); %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % train models in different subspaces %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% traindata=sparse(traindata); testdata=sparse(testdata); for j=1:classifier_num svmmodel(j)=train(traingroup,traindata(:,subspace(j,:)),'-s 3 -c 1 -B 1'); end % trainset=full(trainset); for i=1:testlength % test(i,:)=test(i,:)-mean(test(i,:)); % test(i,:)=test(i,:)/norm(test(i,:)); finalresult=zeros(1,authors); %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % ensembles classifying process %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% for j=1:classifier_num % multi-classifiers construction [predict_label, accuracy] = predict(testgroup(i), testdata(i,subspace(j,:)), svmmodel(j)); finalresult(predict_label+1)=finalresult(predict_label+1)+1; if testgroup(i)==predict_label base_correctnum(fold,j)=base_correctnum(fold,j)+1; end end finalresult if length(find(finalresult==max(finalresult)))>1 % sanme number of votes maxidx=find(finalresult==max(finalresult)); for idx=1:length(maxidx) % compare distance between test and possible class % vectors distance(idx)=dist2(traindata((maxidx(idx)-1)*traintxt+1:maxidx(idx)*traintxt,:),testdata(i,:)); end y=maxidx(find(min(distance))); else [x,y]=max(finalresult); end % [x,y]=max(finalresult); idenidx=y-1; if testgroup(i)==idenidx realauthor = realauthor+1; end fprintf(fid3, '%d%s%d\r\n', testgroup(i),'->',idenidx); end accuracy = (realauthor/(testtxt*authors))*100; accuracys(fold)=accuracy; testdata=[]; traindata=[]; count=count+realauthor; realauthor=0; end maxAccuracy = max(accuracys) totaltest = folds*testtxt*authors count allaccuracy=(count/(totaltest))*100 meanaccuracy=mean(accuracys) base_totalcorrectnum=sum(base_correctnum); base_totalcorrectnum/totaltest fprintf(fid3,'%s%d%s%s\r\n','============================',folds,'folds cross validation','============================'); fprintf(fid3,'%s%d%s%d\r\n','initial dimension:',initial_dimension,'->now dimension:',now_dimension); fprintf(fid3,'%s\t%d\r\n','testset are totally tested in the experiment:', folds*testtxt*authors); fprintf(fid3,'%s\t%d\r\n','realauthors are identified as true:', count); fprintf(fid3,'%s\t%f%s\r\n','allaccuracy of identification:',allaccuracy,'%'); fprintf(fid3,'%s\t%f%s\r\n','meanaccuracy of identification:',meanaccuracy,'%'); toc