multic.m

支持向量域是近几年采用的一种较新的分类器

%MULTIC Make a multi-class classifier
%
%    W = MULTIC(A,V)
%
% Train the (untrained!) one-class classifier V on each of the classes
% in A, and combine it to a multi-class classifier W. If an object is
% rejected by all one-class classifiers, it will be classified
% 'outlier'. If it is accepted by more than one one-class classifier, it
% will be assigned to the class with the highest class posterior.
%
% Frustratingly, it can happen that the class labels are numbers. In
% that case the outlier class will have a 0. For datasets with string
% class labels, the outlier class will be called 'outlier'.
%
%    W = MULTIC(A,{V1 V2 ... VK})
% 
% The trained one-class classifiers V1...VK are combined to a multiclass
% classifier W.
%
% See also dd_normc, myproxm

% Copyright: D.M.J. Tax, [e-mail]
% Faculty EWI, Delft University of Technology
% P.O. Box 5031, 2600 GA Delft, The Netherlands
  
function w = multic(a,w)

[n,k,c] = getsize(a);

% Hmm, labels can be numbers... Make sure that the correct label for the
% outlier class is used.
ll = getlablist(a);
if isa(ll,'double')
%	rejlabel = max(ll)+1;
	rejlabel = 0;
else
	rejlabel = [];
end

% Train the one-class classifiers (if needed) and fix the class labels
% for the multi-class classifier.
for i=1:c
	x = target_class(a,i);
	if length(w)==1
		if istrained(w)
			error('Please supply one untrained one-class mapping.');
		end
		W{i} = x*w*dd_normc;
	else
		if isuntrained(w{i})
			W{i} = x*w{i}*dd_normc;
		else
			W{i} = w{i};
		end
	end
	if isempty(rejlabel)
		W{i} = setlabels(W{i},ll(i,:),1);
	else
		W{i} = setlabels(W{i},[ll(i); rejlabel]);
	end
end

% The final classifier is a fixed combiner:
w = [W{:}]*minc;

return