gauss_dd.m

数据挖掘的工具箱,最新版的,希望对做这方面研究的人有用

%GAUSS_DD Gaussian data description.
% 
%       W = gauss_dd(A,fracrej,r)
% 
% Fit a Gaussian density on dataset A. If requested, the r can be
% given to add some regularization to the estimated covariance matrix:
% sig_new = (1-r)*sig + r*eye(dim). Default r = 0.01!!! (might be
% dangerous!)
%
% This version acutally computes just the Mahalanobis distance to the
% mean. This should avoid underflows at the computation of a real Gaussian
% density (especially problematic in high dimensional spaces).
%
% See also datasets, mappings, dd_roc

% Copyright: D.M.J. Tax, R.P.W. Duin, [e-mail]
% Faculty of Applied Physics, Delft University of Technology
% P.O. Box 5046, 2600 GA Delft, The Netherlands
  
function W = gauss_dd(a,fracrej,r)

if nargin < 3, r = 0.01; end
if nargin < 2 | isempty(fracrej), fracrej = 0.05; end
if nargin < 1 | isempty(a) 
	W = mapping(mfilename,{fracrej});
	W = setname(W,'Gaussian one-class classifier');
	return
end

if ~ismapping(fracrej)           %training

	a = target_class(a);     % only use the target class
	k = size(a,2);

	% Train it:
	[mu,sig] = meancov(+a);
	sig = (1-r)*sig + r*eye(k);

	% Obtain the threshold:
	d = mahaldist(a,mu,sig,-1);
	thr = dd_threshold(d,1-fracrej);

	%and save all useful data:
	W.m = +mu;
	W.s = sig;
	W.threshold = thr;
	W = mapping(mfilename,'trained',W,str2mat('target','outlier'),k,2);
	W = setname(W,'Gaussian one-class classifier');

else                               %testing

	% Extract the data:
	W = getdata(fracrej);
	m = size(a,1);

	% Compute the Mahalanobis distance (to avoid problems in the non-essential
	% normalization factor):
	newout = -[mahaldist(+a,W.m,W.s,-1) repmat(W.threshold,m,1)];

	W = setdat(a,newout,fracrej);
end
return