optim_auc.m

data description toolbox 1.6 单类分类器工具包

function [w,optval] = optim_auc(x,wname,fracrej,range,nrbags,varargin)
%OPTIM_AUC Optimize hyperparameters for an OCC
%
%    W = OPTIM_AUC(X,WNAME,FRACREJ,RANGE,NRBAGS,VARARGIN)
%
% Optimize the AUC-performance of classifier WNAME (string) on dataset
% X. This optimization is over the complexity parameter (this should be
% the third input parameter for the one-class classifier). RANGE should
% contain a vector of parameter values that are tried. For each
% parameter value, NRBAGS-fold crossvalidation is applied and the
% parameter for which the highest AUC is obtained, is used to train the
% final classifier W.
%
% When dataset X does not contain example outlier objects, they
% are generated by gendatout.m.
%
% Example:
% >> a = target_class(gendatb,1);
% >> w = optim_auc(a,'mog_dd',0.1,1:5);
% >> scatterd(a); plotc(w);
%
% See also: scale_range, dd_crossval, gendatout

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

if nargin<5 | isempty(nrbags)
	nrbags = 5;
end
if nargin<4
	range = 1:5;
end
if nargin<3
	fracrej = 0.1;
end
if nargin<2
	wname = 'knndd';
end
if ~exist('x') | isempty(x) 
	% When no inputs are given, we are expected to return an empty
	% mapping:
	if isempty(varargin)
		w = mapping(mfilename,{wname,fracrej,range,nrbags});
	else
		w = mapping(mfilename,{wname,fracrej,range,nrbags,varargin{:}});
	end
	% And give a suitable name:
	w = setname(w,'AUC optim. classifier');
	return
end

% Check some things (there are too many input parameters, and it is very
% easy to forget one, or mix them up):
if ~isa(wname,'char')
	error('Expecting the name (string!) of the classifier.');
end
if length(fracrej)>1
	error('Fracrej should be a scalar.');
end
nrrange = length(range);
if nrrange<2
	error('Expecting a range of param. values.');
end
if ~isa(nrbags,'double') | length(nrbags)>1
	error('Parameter NRBAGS should be a scalar.');
end

% First check if the dataset contains any outlier data, if not then we
% have to generate them!
[xt,xo]=target_class(x);
if size(xo,1)==0
	%fprintf('Generating artificial outliers in "optim_auc"');
	xo = gendatout(x,size(x,1));
end

% run over all the parameter values:
meanauc = zeros(nrrange,1);
for k=1:nrrange

	fprintf('.');
	% Apply the crossvalidation:
	res = zeros(nrbags,1);

	I = nrbags;
	for i=1:nrbags
		% Generate train and test set
		[xtr,xte,I] = dd_crossval(x,I);
		if ~isempty(xo),
			% DXD: sometimes datasets are not completely well-formed, make sure
			% that the identifiers are cellarrays:
			if isa(xte.ident,'double')
				xte.ident = num2cell(xte.ident);
			end
			if isa(xo.ident,'double')
				xo.ident = num2cell(xo.ident);
			end
			xte = [xte; xo];
		else
			error('I require outlier examples for the evaluation of the AUC');
		end
		% Train a classifier:
		if ~isempty(varargin)
			w = feval(wname,xtr,fracrej,range(k),varargin{:});
		else
			w = feval(wname,xtr,fracrej,range(k));
		end
		% Compute the AUC on the leave-out bags
		res(i) = dd_auc(xte*w*dd_roc);
	end
	
	% store the mean auc:
	meanauc(k) = mean(res);
end

% now find which value is the best:
[maxval,kopt] = max(meanauc);

%% So, the final classifier becomes:
w = feval(wname,x,fracrej,range(kopt),varargin{:});
w = setname(w,'AUC optim. classifier');
optval = range(kopt);

return