dlpdd.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

function W = dlpdd(x,nu,usematlab)
%DLPDD Distance Linear Programming Data Description
%
%       W = DLPDD(X,NU)
%
% This linear one-class classifier works directly on the distances, so
% the data is not remapped. The NU parameter gives the fraction of
% error on the target set. It tries to separate the (distance) data as
% best as possible from +infinity.
%
% Although it is more or less assumed that the data is in the positive
% quadrant, you can put other data in as well and see how it may or
% may not work.
%
% See also: lpdd
%
% @inproceedings{Pekalska2002,
%	author = {Pekalska, E. and Tax, D.M.J. and Duin, R.P.W.},
%	title = {One-class {LP} classifier for dissimilarity representations},
%	booktitle = {Advances in Neural Information Processing Systems},
%	year = {2003},
%	pages = {},
%  editor =       {S.~Becker and S.~Thrun and K.~Obermayer},
%  volume =       {15},
%   publisher = {MIT Press: Cambridge, MA}
%}


% Copyright: D. Tax, R.P.W. Duin, [e-mail]
% Faculty of Applied Physics, Delft University of Technology
% P.O. Box 5046, 2600 GA Delft, The netherlands


% first set up the parameters
if nargin < 3, usematlab = 0; end
if nargin < 2 | isempty(nu), nu = 0.05; end
if nargin < 1 | isempty(x) % empty
	W = mapping(mfilename,{nu});
	W = setname(W,'Distance Linear Programming Distance-dd');
	return
end

% training
if ~ismapping(nu)

	% work directly on the distance matrix
	[n,d] = size(x);
	if (n~=d)
		error('I was expecting a square distance matrix!');
	end

	% maybe we have example outliers...
	if isocset(x)
		labx = getoclab(x);
	else
		labx = ones(n,1);
	end
	x = +x; % no dataset please.

	% set up the LP problem:
	C = 1./(n*nu);
	f = [1 -1 zeros(1,n) repmat(C,1,n)]';
	A = [-labx labx repmat(labx,1,n).*x -eye(n)];
	b = -ones(n,1);
	Aeq = [0 0 ones(1,n) zeros(1,n)];
	beq = 1;
	N = 2*n + 2;
	lb = zeros(N,1);
	ub = repmat(inf,N,1);

	% optimize::
	if (exist('lp_solve')>0) & (usematlab==0)
		if ~exist('cplex_init')
			% we can have the lp optimizer:
			e = [0; -ones(n,1)];
			[v,alf] = lp_solve(-f,sparse([Aeq;A]),[beq;b],e,lb,ub);
		else
			% the cplex optimizer:
			lpenv=cplex_init;
			disp = 0;
			[alf,y_upd_,how_upd_,p_lp]=...
			lp_solve(lpenv, f, sparse([Aeq;A]), [beq;b], lb, ub, 1, disp);
		end
	else
		% or the good old Matlab optimizer:
		alf = linprog(f,A,b,Aeq,beq,lb,ub);
	end

	% store the results
	paramalf = alf(3:2+n);
	W.I = find(paramalf>1e-6);
	W.w = paramalf(W.I);
	W.threshold = 1-(alf(1)-alf(2));

	W = mapping(mfilename,'trained',W,str2mat('target','outlier'),d,2);
	W = setname(W,'Distance Linear Programming Distance-data description');

else                               %testing

	% get the data:
	W = getdata(nu);
	m = size(x,1);

	% and here we go:
	D = +x(:,W.I);
	% annoying prtools:
	newout = [-D*W.w repmat(W.threshold,m,1)];

	W = setdat(x,newout,nu);
end
return