source: distools/mclassdc.m

%MCLASSDC Multi-Class Dissimilarity-based Classifier from Two-Class Discriminants
%
%       W = MCLASSDC(D,CLASSF,MODE)
%
% INPUT
%   D       Dissimilarity dataset
%   CLASSF  Untrained dissimilarity-based classifier
%   MODE        Way of handling multi-class problems: 'SINGLE' or 'MULTI' 
%                       (optional; default: 'SINGLE')
%
% OUTPUT
%   W       Combined classifier
%
% DEFAULT
%   MODE = 'SINGLE'
% 
% DESCRIPTION
% For the default MODE = 'SINGLE', the untrained classifier CLASSF is called to
% compute C classifiers between each of the C classes in the dataset D and
% the remaining C-1 classes. The result is stored into the combined
% classifier W. There is no combining rule added. The default rule, MAXC
% might be replaced by adding one, e.g. W = W*MEANC.
%
% For the MODE = 'MULTI', the untrained classifier CLASSF is trained between 
% all pairs of classes as well as between each class and all other classes.
% This total set of C^2 classifiers is combined by MINC.        The use of soft labels
% is supported.
% The pairwise classifiers are trained by using a representation set based
% on the traing objects of the corresponding classes only (just in case of
% crisp labels). This is the only difference with the use of MCLASSC. 
%
% EXAMPLE
% W = MCLASSDC(DISTM(GENDATM(100)),QDC([],0.01),'MULTI')
%
% SEE ALSO
% DATASETS, MAPPINGS, MAXC, MINC

% Copyright: R.P.W. Duin, r.duin@ieee.org
% Faculty EWI, Delft University of Technology 


function varargout = mclassdc(D,classf,mode)

prtrace(mfilename);
if nargin < 3, 
        mode = 'single'; 
end
if nargin < 2, 
        classf = []; 
end
if nargin < 1 | isempty(D)
        w = prmapping(mfilename,{classf,mode});
        return
end
        
if ~ismapping(classf) | ~isuntrained(classf)
        error('Second parameter should be an untrained mapping.')
end

islabtype(D,'crisp','soft');
isvaldset(D,1,2);       % At least 1 object per class and two classes

[m,k,c] = getsize(D);
        
varout = {};
if c == 2
        varargout = prmap(D,classf);
        return
end
lablist = getlablist(D);


switch lower(mode)
        case 'single'
                w = [];
          %     lablist = getlablist(D);
          for i=1:c
                        if islabtype(D,'crisp')
                          mlab = 2 - (getnlab(D) == i);
                          DD = setlabels(D,mlab);
                  elseif islabtype(D,'soft')
                          Dtargets = gettargets(D);
                          targets  = [Dtargets(:,i) 1-Dtargets(:,i)]; 
                          DD                     = prdataset(+D,mlab,targets,'lablist',[1 2]');
                  end
                        if nargout(classf.mapping_file) > 1
                                [v,varo] = prmap(DD,classf);
                                varout   = [varout {varo}];
                        else
                                v = prmap(DD,classf);
                        end
                        w = [w,setlabels(v(:,1),lablist(i,:))];
          end


        case 'multi'
                w = [];
          nlab = getnlab(D);
                [nl,clab,list] = renumlab(getlablist(D),getfeatlab(D));
          for z=1:c
                  lab = lablist(z,:);
                  
                  J1 = find(nlab==z);
                  L1 = find(clab==z);
                  if islabtype(D,'crisp')
                          mlab     = ones(m,1);
                          mlab(J1) = zeros(length(J1),1);
                          DD       = setlabels(D,mlab);
                  else
                          problab  = gettargets(D);
                          mlab     = [problab(:,i1) sum(problab,2)-problab(:,z)];
                          DD       = settargets(D,mlab,[1 2]');
                  end           
                  I1 = setdiff([1:c],z);
                        if nargout(classf.mapping_file) > 1
                                [v,varo] = prmap(DD,classf);
                                varout   = [varout {varo}];
                        else
                                v = prmap(DD,classf);
                        end
                        w = [w,setlabels(v(:,1),lab)];

                  for t = I1 
                          if islabtype(D,'crisp')
                                  J2 = find(nlab==t);
                                  L2 = find(clab==t);
                                  v  = featsel(k,[L1;L2])*(DD([J1;J2],[L1;L2])*classf);
                                        %disp([z,t]);
                                        %parsc(v);
                          else
                                  mlab2 = problab(:,[z t]);
                                  v     = setlabels(DD,mlab2)*classf;
                          end
                          w = [w,setlabels(v(:,1),lab)];
                  end
          end
          w = minc(w);
                
        otherwise
          error('Unknown mode')
end

w = setname(w,getname(classf));
w = setsize(w,[k,c]);
w = setcost(w,D);
varargout = {w varout};
return