[Home]

##### TWSVC

Matlab code for twin bounded support vector clustering (TBSVC) and least squares twin bounded support vector clustering (LSTBSVC).

##### Reference

Lan Bai, Yuan-Hai Shao*, Zhen Wang, Chun-Na Li. Clustering by twin support vector machine and least square twin support vector classifier with uniform output coding[J]. Submitted. 2018

##### Main Function

pY= TBSVC(X,Y,c1,c2) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % TBSVC: % % pY= TBSVC(X,Y,c1,c2) % % Input: % X: Data matrix. (nolinear version is obtained by K(X,X)) % Y: First initial labels of X (can be given randomly or by NN-graph). % Parameters - c1,c2. The fields in options that can be set: % c1, c2: (0,inf) Paramter to tune the weight. % % Output: % pY: The prediction of X. % Examples: % X=rand(50,10); % Y = randint(50,1,[1,5]); % c1=1;c2=1;; % pY= TBSVC(X,Y,c1,c2) % % Version 1.0 --Apr/2018 % % Written by Zhen Wang (wangzhen@imu.edu.cn) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% tol=0.001; num=max(Y); totalu=zeros(1+size(X,2),num); maxite=30; flag=0; outite=0; pY=Y; while flag==0 && outitetol && ite<30 ite=ite+1; u=u0; e1=ones(m1,1); e2=ones(m2,1); G=[inputB,e2]; D=diag(sign(G*u))*G; H=[inputA,e1]; K=(H'*H+diag([c1*ones(n,1);0]))\D'; %gamma=quadprog(); gamma=qpSOR(D*K,0.7,c2,0.05); % using SOR method, one may use the above to solve it u0=K*gamma; som=norm(u-u0); end totalu(:,i)=u0; end for i=1:num totalu(:,i)=totalu(:,i)/norm(totalu(1:n,i)); end [~,pY]=min(abs([X,ones(size(X,1),1)]*totalu),[],2); if getAC(pY,tY)>0.9999 % check if there are some changes flag=1; end end end
pY= LSTBSVC(X,Y,c1,c2) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % LSTBSVC: % % pY= LSTBSVC(X,Y,c1,c2) % % Input: % X: Data matrix. (nolinear version is obtained by K(X,X)) % Y: First initial labels of X (can be given randomly or by NN-graph). % Parameters - c1,c2. The fields in options that can be set: % c1, c2: (0,inf) Paramter to tune the weight. % % Output: % pY: The prediction of X. % Examples: % X=rand(50,10); % Y = randint(50,1,[1,5]); % c1=1;c2=1;; % pY= LSTBSVC(X,Y,c1,c2) % % Version 1.0 --Apr/2018 % % Written by Zhen Wang (wangzhen@imu.edu.cn) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% num=max(Y); totalu=zeros(1+size(X,2),num); maxite=30; flag=0; outite=0; pY=Y; m=size(X,1); while flag==0 && outite< maxite outite=outite+1; tY=pY; for i=1:num inputA=X(tY==i,:); inputB=X(tY~=i,:); [m1,n]=size(inputA); m2=size(inputB,1); H=[inputA,ones(m1,1)]; G=[inputB,ones(m2,1)]; totalu(:,i)=(H'*H+G'*G*c2+c1*diag([ones(n,1);0]))\(G'*ones(m2,1))*c2; end % Prediction by code for i=1:m Dis=zeros(2*num,1); for j=1:num val=X(i,:)*totalu(1:n,j)+totalu(n+1,j); Dis(j,1)=abs(val)/norm(totalu(1:n,j)); Dis(num+j,1)=abs(val-1)/norm(totalu(1:n,j)); end code=Code(Dis,num); sumcode=sum(code,2); ind=find(sumcode==max(sumcode)); if length(ind)==1 pY(i,1)=ind; else normu=zeros(1,num); for t=1:num normu(1,t)=norm(totalu(:,t)); end [~,pY(i,1)]=min(abs([X(i,:),1]*totalu)./normu,[],2); end end if getAC(pY,tY)>0.9999 flag=1; end end end %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Additional functions %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% function u=FirstStep(A) % compute: min ||Aw+be||, s.t. ||w||=1. % u=[w;b] m=size(A,1); H=A'*(1/m*ones(m,m)-eye(m))*A; [V,D]=eig(H); [tmp,n]=min(abs(diag(D))); w=V(:,n); b=-1/m*sum(A,1)*w; u=[w;b]; end function code=Code(Dis,num) % Dis (2*num,1): |f1|, |f2|, |f3|, |f1-1|, |f2-1|, |f3-1| % w(k), c(k+1) % c1 1 -1 -1 -1 1 1 1 -1 0 -1 1 0 % c2 -1 1 -1 1 -1 1 -1 1 0 1 -1 0 % c3 ... code=zeros(num,4*num); for i=1:num for j=1:num if Dis(j)<=Dis(j+num) if j==i code(i,j)=1; else code(i,j)=-1; end else if j==i code(i,j+num)=-1; else code(i,j+num)=1; end end end [~,ind]=min(Dis(1:num)); if ind==i if code(i,i)==1 code(i,2*num+i)=2; else code(i,2*num+i)=0; end else code(i,2*num+ind)=-1; if code(i,ind)==-1 code(i,2*num+ind)=-2; else code(i,2*num+ind)=0; end end val=sum(code(i,1:2*num))/num; code(i,1:2*num)=zeros(1,2*num); code(i,1)=val; end end
##### Contacts

Any question or advice please email to wangzhen@imu.edu.cn.

• Last updated: Apr 21, 2018