%FEATSELB Trainable mapping for backward feature selection
%
% [W,R] = FEATSELB(A,CRIT,K,T)
% [W,R] = A*FEATSELB([],CRIT,K,T)
% [W,R] = A*FEATSELB(CRIT,K,T)
% [W,R] = FEATSELB(A,CRIT,K,N)
% [W,R] = A*FEATSELB([],CRIT,K,N)
% [W,R] = A*FEATSELB(CRIT,K,N)
%
% INPUT
% A Dataset
% CRIT String name of the criterion or untrained mapping
% (optional; default: 'NN', i.e. 1-Nearest Neighbor error)
% K Number of features to select
% (optional; default: return optimally ordered set of all features)
% T Tuning set (optional)
% N Number of cross-validations
%
% OUTPUT
% W Output feature selection mapping
% R Matrix with step-by-step results of the selection
%
% DESCRIPTION
% Backward selection of K features using the dataset A. CRIT sets the
% criterion used by the feature evaluation routine FEATEVAL. If the
% dataset T is given, it is used as test set for FEATEVAL. Alternatvely a
% a number of cross-validation N may be supplied. For K = 0, the optimal
% feature set (corresponding to the maximum value of FEATEVAL) is returned.
% The result W can be used for selecting features by B*W. In this case,
% features are ranked optimally.
% The selected features are stored in W.DATA and can be found by +W.
% In R, the search is reported step by step as:
%
% R(:,1) : number of features
% R(:,2) : criterion value
% R(:,3) : added / deleted feature
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% MAPPINGS, DATASETS, FEATEVAL, FEATSELLR, FEATSEL,
% FEATSELO, FEATSELF, FEATSELI, FEATSELP, FEATSELM
% Copyright: R.P.W. Duin, r.p.w.duin@37steps.com
% Faculty EWI, Delft University of Technology
% P.O. Box 5031, 2600 GA Delft, The Netherlands
% $Id: featselb.m,v 1.6 2008/07/03 09:08:43 duin Exp $
function [w,r] = featselb(varargin)
varargin = shiftargin(varargin,{'char','prmapping'});
argin = setdefaults(varargin,[],'NN',0,[],[]);
if mapping_task(argin,'definition')
w = define_mapping(argin,'untrained','Backward FeatSel');
return
end
[a,crit,ksel,t,fid] = deal(argin{:});
[w,r] = featsellr(a,crit,ksel,0,1,t,fid);
%DXD This is a patch: when the number of features has to be
%optimized, and all features seem useful, when the list of
%features is not reshuffled to reflect the relative importance of
%the features:
% (Obviously, this should be fixed in featsellr, but I don't
% understand what is happening in there)
dim = size(a,2);
if (ksel==0) & (length(getdata(w))==dim)
rr = -r(:,3); rr(1) = [];
rr = [setdiff((1:dim)',rr) rr(end:-1:1)'];
w = setdata(w,rr);
end
return