%KERNELC Arbitrary kernel/dissimilarity based classifier
%
% W = KERNELC(A,KERNEL,CLASSF)
% W = A*KERNELC([],KERNEL,CLASSF)
% W = A*KERNELC(KERNEL,CLASSF)
%
% INPUT
% A Dateset used for training
% KERNEL - untrained mapping to compute kernel by A*(A*KERNEL) for
% training CLASSF or B*(A*KERNEL) for testing with dataset B, or
% - trained mapping to compute a kernel A*KERNEL for training
% CLASSF or B*KERNEL for testing with a dataset B
% KERNEL should be a functions like PROXM, KERNELM or USERKERNEL.
% Default: reduced dissimilarity representation:
% KERNELM([],[],'random',0.1,100);
% CLASSF Classifier used in kernel space, default LOGLC.
%
% OUTPUT
% W Resulting, trained classifier
%
% DESCRIPTION
% This routine defines a classifier W in the input feature space based
% on a kernel or dissimilarity representation defined by KERNEL and a
% classifier CLASSF to be trained in the kernel space.
%
% In case KERNEL is defined by V = KERNELM( ... ) this routine
% is identical to W = A*(V*CLASSF), Note that if KERNEL is a mapping, it
% may be trained as well as untrained. In the latter case A is used to
% build the kernel space as well as to optimize the classifier (like in SVC).
%
% EXAMPLE
% a = gendatb([100 100]); % training set of 200 objects
% r = gendatb([10 10]); % representation set of 20 objects
% v = proxm(r,'p',3); % compute kernel
% w = kernelc(a,v,fisherc) % compute classifier
% scatterd(a); % scatterplot of trainingset
% hold on; scatterd(r,'ko'); % add representation set to scatterplot
% plotc(w); % plot classifier
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% DATASETS, MAPPINGS, KERNELM, PROXM, USERKERNEL
% 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 = kernelc(varargin)
argin = shiftargin(varargin,'prmapping');
argin = setdefaults(argin,[],kernelm([],[],'random',0.7,100),fisherc);
if mapping_task(argin,'definition')
w = define_mapping(argin,'untrained');
else % training
[a,kernel,classf] = deal(argin{:});
islabtype(a,'crisp');
isvaldfile(a,1,2); % at least 1 object per class, 2 classes
a = testdatasize(a,'objects');
isuntrained(classf);
ismapping(kernel);
if isuntrained(kernel)
kernel = a*kernel;
end
K = a*kernel;
w = kernel*(K*classf);
end
w = setname(w,'Kernel Classifier');
return