dd_roc.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 e = dd_roc(a,w)
%DD_ROC Receiver Operating Characteristic curve 
%
%        E = DD_ROC(A,W)
%        E = DD_ROC(A*W)
%        E = A*W*DD_ROC
%
% Find for a (data description) method W the Receiver Operating
% Characteristic curve over dataset A.  The results are returned in E.
% The first column gives the fraction of target objects rejected, the
% second column the fraction of outlier objects accepted.
%
% NOTE: people typically use this ROC definition: false positive FP
% (outlier accepted) on the x-axis, and true positive TP (target
% accepted) on the y-axis. You can retrieve that by using:
%   newe = [e(:,2) 1-e(:,1)]
% I choose to define E consistent, i.e. both numbers indicate
% 'errors'. In the routines plotroc and dd_auc the variable E is
% automatically recomputed to get the 'standard' plots and AUC values.
%
% See also: plotroc, dd_auc, dd_error, simpleroc.
%
%@article{Metz1978,
%	author = {Metz, C.E.},
%	title = {Basic principles of {ROC} analysis},
%	journal = {Seminars in Nuclear Medicine},
%	year = {1978},
%	volume = {VIII},
%	number = {4},
%	month = {October}
%}


% 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

% Use the same setup as testc

% When no input arguments are given, return an empty mapping
if nargin==0
	
	e = mapping(mfilename,'fixed');

elseif nargin == 1

	% Now we should have a mapped dataset, so the real work is done!

	% for evaluation, we need both target and outlier objects:
	dataset(a);
	[It,Io] = find_target(a);
	if isempty(It)
		error('Dataset A does not contain target objects');
	end
	if isempty(Io)
		error('Dataset A does not contain outlier objects');
	end

	% get the labels of A:
	truelab = zeros(size(a,1),1);
	truelab(It) = 1;

	% check if we have sane results:
	if ~all(isfinite(+a))
		error('Some strange classifier outputs: can you check your classifier?');
	end

	% first find out where the output for the target objects are stored:
	tcolumn = strmatch('target ',getfeatlab(a));
	if tcolumn~=1
		% then we are probably using 'normal' prtools classifiers, and in
		% that case, the outputs should be normalized
		if abs(sum(sum(+a)) - size(a,1)) > 1e-9
			error('Are the classifier outputs normalized?');
	  end
	end
	% and now extract the required column 'resemblance to target set':
	a = +a(:,tcolumn);

	% now the real computation is done:
	e = simpleroc(a,truelab);

else

	% Separate mapping and dataset are given, so we have to map the data
	% first:
	ismapping(w);
	istrained(w);

	e = feval(mfilename,a*w);

end

return