source: prextra/ecoc.m @ 89

%ECOC Error-correcting output code
% 
%      W = ECOC(A,BASECL,CM,RULE)
%
% INPUT
%    A       Dataset
%    BASECL  Untrained base classifier
%    CM      Coding matrix
%    RULE    Combination rule
%
% OUTPUT
%    W       ECOC  classifier
%
% DESCRIPTION
% Computation of the classifier using Error-correcting output codes to
% create a multi-class classifier from a base classifier/dichotomizer.
% The classes in dataset A are relabeled according to the coding matrix
% CM (containing +1, -1 or 0). The default coding matrix is a 1-vs-rest
% coding. For a three-class problem this becomes:
%     CM = [ 1 -1 -1;
%           -1  1 -1;
%           -1 -1  1];
%
% For the evaluation of objects, the outputs of the dichotomizers are
% combined using the combination rule RULE. Currently the following
% rules are implemented:
%     'none'     take the maximum output (this only works for 1-vs-all
%                coding matrices)
%     'hamming'  standard Hamming distance between the discretized
%                classifier outcomes and the coding matrix. When a zero
%                entry in the coding matrix appears, this entry is not
%                used in the distance computation
%
% NOTE: the order of the classes as they are used in the coding matrix
% is determined by the lablist of A.
%
% SEE ALSO
%    MCLASSC, MAXC

% Copyright: D.M.J. Tax, D.M.J.Tax@prtools.org
% Faculty EWI, Delft University of Technology
% P.O. Box 5031, 2600 GA Delft, The Netherlands
  
function W = ecoc(a,basecl,CM,rule)

if nargin < 4 | isempty(rule), rule = 'hamming'; end
if nargin < 3 | isempty(CM), CM = []; end
if nargin < 2 | isempty(basecl), basecl = fisherc; end
if nargin < 1 | isempty(a) 
        W = mapping(mfilename,{w,L,rule});
        W = setname(W,'ECOC');
        return
end

if ~ismapping(basecl) | ~istrained(basecl)   %training

        [n,k,c] = getsize(a);
        nlab = getnlab(a);
        % define the coding matrix if not given:
        if isempty(CM)
                CM = repmat(-1,c,c) + 2*eye(c);
        else
                % check if the coding matrix is OK
                if size(CM,1)~=c
                        error('Coding matrix does not have correct number of rows.');
                end
                if (max(CM(:))>1)|(min(CM(:))<-1)
                        error('Entries in the coding matrix should be -1 or +1.');
                end
        end
        l = size(CM,2);
   w = cell(1,l);

        % Train it:
        for i=1:l
                % create a new dataset
                lab = nlab;
                for j=1:c
                        lab(lab==j) = CM(j,i);
                end
                % should I remove the objects with 0 label?
                b = setlabels(a,lab);
                % train the base dichotomizer:
                w{i} = b*basecl;
        end

        %and save all useful data:
        W.CM = CM;
        W.w = w;
        W.rule = rule;
        W = mapping(mfilename,'trained',W,getlablist(a),k,c);
        W = setname(W,'ECOC');

else                               %testing

        % Extract the data:
        W = getdata(basecl);
        m = size(a,1);
        [c,l] = size(W.CM);

        % Apply all the base classifiers:
        z = zeros(m,l);
        for i=1:l
                tmp = a*W.w{i};
                id = findfeatlab(tmp,1);
                z(:,i) = +tmp(:,id);
        end
        z = 2*z-1;  %DXD classifier output should be between -1 and +1!!!
%[W.CM; z]
        % and apply the combination rule:
        out = zeros(m,c);
        switch W.rule
        case 'none'
                if size(z,2)~=c
                        error('You have to combine when you don''t do one-vs-rest.');
                end
                out = z;
        case 'hamming'
                for i=1:m
                        zz = sign(z(i,:)); % discrete output
                        % hamming distance of object i to all code words:
                        % remove entries for which the coding matrix has a 0:
                        CM0 = (W.CM~=0);
                        d = sum(CM0.*abs(repmat(zz,c,1)-W.CM),2);
%keyboard
                        [md,Icl] = min(d);
                        % class Icl won!
                        out(i,Icl) = 1;
                end
        otherwise
                error('This combination rule is not implemented yet.');
        end

        % Store the outputs
        W = setdat(a,out,basecl);
end
return