mog_dd.m

data description toolbox 1.6 单类分类器工具包

%MOG_DD Mixture of Gaussians data description
%
%   W = MOG_DD(A,FRACREJ,[N1 N2],CTYPE,REG,NUMITERS)
%
% Train a Mixture of Gaussians model on data A, using N1 clusters to
% model the target class, and N2 clusters for the outlier data. The
% position, size and priors of each of the clusters is optimized using
% the EM algorithm. The shape of the clusters can be chosen by setting
% CTYPE:
% CTYPE = 'sphr':  spherical clusters (only diagonal elements with
%                  equal values in the cov. matrices)
%         'diag':  elliptical and aligned clusters (only diagonal
%                  elements)
%         'full':  arbitrary shaped clusters (full covariance matrices)
%
% When no outlier data is available, you can choose to use
% N2=[], or N2=0. To ensure that the classifiers obtains a closed
% boundary around the target class, one extra cluster is added to the
% outlier clusters, that will model the uniform outlier distribution.
% When you only supply N1, a fixed threshold will be used and no extra
% outlier cluster is added.
%
% The EM algorithm was inspired by NetLab (by Bishop), but it has been
% changed significantly after that.
% Optionally, you can set a regularization for the covariance matrices,
% by giving REG, or the maximum number of iterations by NUMITERS.
%
% Default: N1=5, CTYPE='full', REG=0.01, NUMITERS=25
%
% See also: mog_init, mog_update, mog_P, mogEMupdate, mogEMextend

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

function w = mog_dd(a,fracrej,n,ctype,reg,numiters)

if (nargin<6)
	numiters = 25;
end
if (nargin<5)
	reg = 0.01;
end
if (nargin<4)
	ctype = 'full';
end
if (nargin<3)
	n = 5;
end
if (nargin<2)
	fracrej = 0.05;
end
if length(n)>1  % then we want to train using outlier data
	if (n(2)<0), n(2)=0; end; % using at least 0 outlier clusters
end
if (nargin<1)|isempty(a)     % Define the untrained mapping.
	w = mapping(mfilename,{fracrej,n,ctype,reg,numiters});
	if (length(n)>1) & (n(2)>0)  % we model also the outliers
		w = setname(w,sprintf('MoG (%d+%d)',n(1),n(2)));
	else
		% we only describe the target class
		w = setname(w,sprintf('MoG (%d)',n(1)));
	end
	return
end

if isa(fracrej,'double')           %training

	% Setup some variables:
	W = [];
	[m,k] = size(a);
	[at,ao] = target_class(a);

	% Detect which type of covariance matrix we have:
	switch ctype
	case {'sphr' 'diag' 'full'}
		W.covtype = ctype;
	otherwise
		error('This covariance structure is not known');
	end

	% First do the target class
	[W.mt, W.ict, W.pt] = mog_init(at, n(1), W.covtype);
	% then do the outlier class
	if (length(n)>1)
		if isempty(ao)
			% When no outlier data is avaiable, all clusters will have
			% centers on the mean of the data
			[W.mo, W.ico, W.po] = mog_init(mean(at),n(2),W.covtype,cov(+at));
		else
			% When some outlier data is available, only the first cluster
			% will be fixed to the mean of the data, the rest will be
			% fitted
			[W.mo, W.ico, W.po] = mog_init(ao,n(2),W.covtype,cov(+a));
		end
	end
	
	% and not to forget to store:
	W.threshold = 0;
	W.reg = reg;
	W.fracrej = fracrej;

	% and store it in the mapping:
	w = mapping(mfilename,'trained',W,str2mat('target','outlier'),k,2);
	% also give it a suitable name
	if (length(n)>1) & (n(2)>0)  % we model also the outliers
		w = setname(w,sprintf('MoG (%d+%d)',n(1),n(2)));
	else
		% we only describe the target class
		w = setname(w,sprintf('MoG (%d)',n(1)));
	end
	
	% Now start training this mapping:
	w = mog_update(w,a,numiters);
	w = mog_threshold(w,at,fracrej);

else                               %testing

	w = getdata(fracrej);  % unpack
	m = size(a,1);

	%compute the output, depending if the outlier class is also modelled:
	out = sum(mog_P(+a,w.covtype,w.mt,w.ict,w.pt),2);
	% when we have the outlier class modelled:
	if isfield(w,'mo')
		% Do it nicely using Bayes rule:
		Po = sum(mog_P(+a,w.covtype,w.mo,w.ico,w.po),2) + 10*eps;
		s = warning('off');
			out = out./Po;
		warning(s);
	end
	newout = [out repmat(w.threshold,m,1)];

	% Store the density:
	w = setdat(a,newout,fracrej);
	w = setfeatdom(w,{[0 inf] [0 inf]});
end

return