%mi-svm
%data.mat is the matfile of the attributes from f1-f166;
%Y.mat the frist column is every instance's bag label and the second column
%is the instances number sequence of every bags.
load data.mat;
load Ylabel.mat
patternlabel1=zeros(size(Ylabel,1),1);
patternlabel2=Ylabel(:,1);
while(patternlabel1~=patternlabel2) %need improve,"=="
%caller C-svm programme,solve the W,b,and solution the pattern label F.
%method select the best parameter by the libsvm, and then use the best
%parameter in this code. so we can saving the time.
C=input('please input the best C');
sigma=input('please input the best sigma');
ker = struct('type','gauss','width',sigma);
X=data';
Y=patternlabel2';
patternlabel1=patternlabel2;
options=optimset;
options.largeScale='on';
options.MaxIter=4000;
n=length(Y);
H1=Y'*Y;
H2=kernel(ker,X,X);
H=(Y'*Y).*kernel(ker,X,X);
f=-ones(n,1);
A=[];
b=[];
Aeq=Y;
beq=0;
lb=zeros(n,1);
ub=C*ones(n,1);
a0=ones(n,1);
[a,fval,eXitflag,output,lambda] = quadprog(H,f,A,b,Aeq,beq,lb,ub,a0,options);
tmp=(a'.*Y)*kernel(ker,X,X);
b=Y-tmp;
b=mean(b);
tmp=(a'.*Y)*kernel(ker,X,X);
F=tmp+b;
y=sign(F);
i=1;
while(Ylabel(i,1)==1)
if(Ylabel(i,2)>=Ylabel(i+1,2))
m=Ylabel(i,2);
bagf=F(1,i-m+1:i);
bagy=y(1,i-m+1:i);
if (sum(bagy)+m)/2==0
istar=find(max(bagf));
y(1,i-m+istar)=1;
end
end
i=i+1;
end
patternlabel2=y';
end
%find the best results W and b,then use the best result to solve the test
%data and decise the test label.
load testdata.mat
load testbaglabel.mat
testdata=testdata';
testtem=(a'.*Y)*kernel(ker,X,testdata);
testy=sign(testtem+b);
N=length(testbaglabel);
j=0;
baglabel=0;
for h=1:(N-1)
if testbaglabel(h)>=testbaglabel(h+1)
j=j+1;
numbag=testbaglabel(h);
testbagy=testy(1,h-numbag+1:h);
label=max(testbagy);
baglabel(j)=label;% this is the label of bag;
labelpattern(h-numbag+1:h,1)=label*ones(numbag,1);%this is the label of every instance.
end
ends
totalbag=j;
baglabel;
testy;
%the follow is the code of count the wright rate and discuss the
%performance of the algorithm.
