source: prextra/emc.m @ 5

%EMC EM Classifier using semi-supervised data
%
%   W = EMC(A,B,CLASSF,LABTYPE,FID)
%   W = A*EMC([],B,CLASSF,LABTYPE,FID)
%
% INPUT
%   A       Labeled dataset used for training
%   B       Additional unlabeled dataset
%   CLASSF  Untrained classifier (default QDC)
%   LABTYPE Label type to be used (crisp (default) or soft)
%   FID     File ID to write progress to (default [], see PRPROGRESS)
%
% OUTPUT
%   W      Trained classifier
%
% DESCRIPTION
% Using the EM algorithm the classifier CLASSF is used iteratively
% on the joint dataset [A;B]. In each step the labels of A are reset
% to their initial values. Initial labels in B are neglected.
% Labels of LABTYPE 'soft' are not supported by all classifiers. 
%
% SEE ALSO
% DATASETS, MAPPINGS, EMCLUST, PRPROGRESS

% Copyright: R.P.W. Duin, r.p.w.duin@prtools.org
% Faculty EWI, Delft University of Technology
% P.O. Box 5031, 2600 GA Delft, The Netherlands

function w = emc(a,b,classf,labtype,fid)
        if nargin < 5, fid = []; end
        if nargin < 4 | isempty(labtype), labtype = 'crisp'; end
        if nargin < 3 | isempty(classf), classf = qdc; end      
        if nargin < 2, b = []; end
        if nargin < 1 | isempty(a)
                w = mapping(mfilename,'untrained',{b,classf,labtype,fid});
                w = setname(w,'EM CLassifier');
                return
        end

        islabtype(a,'crisp','soft');
        isvaldset(a,1,2); % at least 2 object per class, 2 classes
        if isempty(b)
                w = a*classf;
                return
        end
        if size(a,2) ~= size(b,2)
                error('Datasets should have same number of features')
        end
        
        c = getsize(a,3);
        epsilon = 1e-6;
        change = 1;
        nlab = getnlab(a);
        lablist = getlablist(a);
        a = setlabels(a,nlab); 
        a = setlabtype(a,labtype);
        switch labtype
                case 'crisp'
                        lab = zeros(size(b,1),1);
                case 'soft'
                        lab = zeros(size(b,1),c);
        end
        b = dataset(+b);
        w = a*classf;
        
        prprogress(fid,'\nem_classifier optimization\n')                
        while change > epsilon
                d = b*w;
                switch labtype
                        case 'crisp'
                                labb = d*labeld;
                                change = mean(lab ~= labb); 
                        case 'soft'
                                labb = d*classc;
                                change = mean(mean((+(labb-lab)).^2));
                        otherwise
                                error('Wrong LABTYPE given')
                end
                lab = labb;
                b = setlabtype(b,labtype,lab);
                c = [a; b];
                w = c*classf;
                prprogress(fid,'  change = %d\n', change)
        end
        
        J = getlabels(w);
        w = setlabels(w,lablist(J,:));