svmroc.m

基于核分析的多类分类器

function [xsup,w,w0,pos,timeps,alpha,matriceind]=svmroc(x,y,C,kppv,margin,lambda,kernel,kerneloption,verbose,span,qpsize,kkttol,matriceind,alphainit);

% USAGE
% 
% [xsup,w,w0,pos,timeps,alpha,matriceind]=svmroc(x,y,C,kppv,margin,lambda,kernel,kerneloption,verbose,span,qpsize,matriceind)
%
%
% SVM ROC Optimizer that can handle LS problem and large neighboorhood. This
% algorithm uses a decomposition procedure.
%
%  x y          the learning data and labels
%  C            penalization parameters
%  kppv         the number of neighboor to consider. choose kppv=inf for
%               genuine ROC-SVM with no approx.
%  margin       the margin for ranking
%  lambda       conditioning parameter for the qp problem e.g 1e-7
%  kernel       the kernel type e.g 'gaussian' or 'poly'
%  kerneloption kernel parameters
%  verbose      verbosity of the algo 0 or 1
%  span         type of semi parametric function e.g 1
%  qpsize       size of qp algorithm
%
%
% Outputs as usual for SVM except that xsup is a cell containing
% the couple of positive and negative support vectors.
% 
% see also svmrocval
%
% 

% 30/07/2004 A. Rakotomamonjy

if nargin< 14
    alphainit=[];
end;
if nargin< 13
    matriceind=[];
end;
if nargin< 12
    kkttol=1e-4;
end;

if size(alphainit,1)~=size(matriceind,1)
    error('matriceind and alphainit initialization size mismatch...');
end;

iteration=0;
difftol=1e-7;
chunksize=qpsize;

indpos=find(y==1);
indneg=find(y==-1);
nbpos=length(indpos);
nbneg=length(indneg);

timeps=0;


%-------------------------------------------------------------------------
%
% dist in feature space of each positive example to negative examples
%
% the idea here is to select a subset of couple for
% optimizing the ranking
%
% 
% CASE 1 : select only the k-nearest negative neighbor of positive examples
if isempty(matriceind);
    if kppv~=inf
        for i=1:nbpos+nbneg
            norme2(i)=svmkernel(x(i,:),kernel,kerneloption);
        end;
        
        matriceindneg=[];
        % select only the k-nearest positive neighbor of negative xamples
        for i=1:nbneg
            aux1=svmkernel(x(indneg(i),:),kernel,kerneloption,x(indpos,:));
            dist=norme2(indneg(i))*ones(1,nbpos) + norme2(indpos) - 2*aux1 ;
            [aux,indicesorted]=sort(dist');
            minim=min(length(indicesorted),kppv);
            matriceindneg=[matriceindneg; i*ones(minim,1) indicesorted(1:minim)];
        end;
        vect=unique(matriceindneg(:,2)); 
        % process only these couples of nn of these positive samples.
        dist=[];
        matriceind=[];
        for i=1:length(vect)
            aux1=svmkernel(x(indpos(vect(i)),:),kernel,kerneloption,x(indneg,:));
            dist=norme2(indpos(vect(i)))*ones(1,nbneg) + norme2(indneg) - 2*aux1 ;
            [aux,indicesorted]=sort(dist');
            minim=min(length(indicesorted),kppv);
            matriceind=[matriceind; vect(i)*ones(minim,1) indicesorted(1:minim)];
        end;
        
    else   % Select all couples
        %         k=1;
        %         for i=1:nbneg
        %             for j=1:nbpos
        %                 matriceind(k,:)=[j i];
        %                 k=k+1;    
        %             end;
        %         end;
        
        [aux1 aux2]  = meshgrid(1:nbpos,1:nbneg); 
        [nn1,nn2]= size(aux1); 
        matriceind = [reshape(aux1 ,nn1*nn2,1) reshape(aux2 ,nn1*nn2,1)]; 
        
    end;
end;
taille=length(matriceind);
%--------Matrice stocke la liste des couples de points