%FEATSELM Feature selection map
%
% [W,R] = FEATSELM(A,CRIT,METHOD,K,T,PAR1,...)
%
% INPUT
% A Training dataset
% CRIT Name of criterion: 'in-in', 'maha-s', 'NN' or others
% (see FEATEVAL) or an untrained classifier V (default: 'NN')
% METHOD - 'forward' : selection by featself (default)
% - 'float' : selection by featselp
% - 'backward': selection by featselb
% - 'b&b' : branch and bound selection by featselo
% - 'ind' : individual
% - 'lr' : plus-l-takeaway-r selection by featsellr
% - 'sparse' : use sparse untrained classifier CRIT
% K Desired number of features (default: K = 0, return optimal set)
% T Tuning set to be used in FEATEVAL (optional)
% PAR1,.. Optional parameters:
% - L,R : for 'lr' (default: L = 1, R = 0)
%
% OUTPUT
% W Feature selection mapping
% R Matrix with step by step results
%
% DESCRIPTION
% Computation of a mapping W selecting K features. This routines offers a
% central interface to all other feature selection methods. W can be used
% for selecting features in a dataset B using B*W.
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% MAPPINGS, DATASETS, FEATEVAL, FEATSELO, FEATSELB, FEATSELI,
% FEATSELP, FEATSELF, FEATSELLR
% 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
function [w,res] = featselm(a,crit,arg3,ksel,t,par1,par2)
if (nargin < 2 | isempty(crit))
prwarning(2,'criterion not specified, assuming NN');
crit = 'NN';
end
if (nargin < 3 | isempty(arg3))
prwarning(2,'method not specified, assuming forward');
arg3 = 'forward';
end
if (nargin < 4)
ksel = [];
end
if (nargin < 5)
prwarning(3,'no tuning set supplied (risk of overfit)');
t = [];
end
if (nargin < 6), par1 = []; end;
if (nargin < 7), par2 = []; end;
% If no arguments are supplied, return an untrained mapping.
if (nargin == 0) | (isempty(a))
w = prmapping('featselm',{crit,arg3,ksel,t,par1,par2});
w = setname(w,'Feature Selection');
return
end
a = testdatasize(a);
[m,k] = size(a);
if (isstr(arg3))
method = arg3; % If the third argument is a string,
switch (method) % it specifies the method to use.
case {'forward','featself'}
[w,res] = featself(a,crit,ksel,t);
case {'float','featselp'}
[w,res] = featselp(a,crit,ksel,t);
case {'backward','featselb'}
[w,res] = featselb(a,crit,ksel,t);
case {'b&b','featselo'}
[w,res] = featselo(a,crit,ksel,t);
case {'ind','featseli'}
[w,res] = featseli(a,crit,ksel,t);
case {'lr','featsellr'}
[w,res] = featsellr(a,crit,ksel,par1,par2,t);
case {'sparse'}
v = a*crit;
if isaffine(v)
v = getdata(v,'rot');
w = featsel(size(a,2),find(v(:,1) == 0));
else
v = getdata(v,'beta');
end
w = featsel(size(a,2),find(v(:,1) ~= 0));
otherwise
error('Unknown method specified.')
end
elseif (ismapping(arg3))
w = arg3; % If the third argument is a mapping,
isuntrained(w); % assert it is untrained and train
[w,res] = feval(mfilename,a,crit,w.mapping_file,ksel,t,par1,par2);
else
error('Illegal method specified.')
end
return