%RSSCC Trainable random subspace combining classifier
%
% W = RSSCC(A,CLASSF,NFEAT,NCLASSF)
% W = A*RSSCC([],CLASSF,NFEAT,NCLASSF)
% W = A*RSSCC(CLASSF,NFEAT,NCLASSF)
%
% INPUT
% A Dataset
% CLASSF Untrained base classifier
% NFEAT Number of features for training CLASSF
% NCLASSF Number of base classifiers
%
% OUTPUT
% W Combined classifer
%
% DESCRIPTION
% This procedure computes a combined classifier consisting out of NCLASSF
% base classifiers, each trained by a random set of NFEAT features of A.
% W is just the set of base classifiers and still needs a combiner, e.g.
% use W*MAXC or W*VOTEC.
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% DATASETS, MAPPINGS, PARALLEL
function w = rsscc(varargin)
mapname = 'rsscc';
argin = shiftargin(varargin,'prmapping');
argin = setdefaults(argin,[],nmc,[],[]);
if mapping_task(argin,'definition')
w = define_mapping(argin,'untrained',mapname);
elseif mapping_task(argin,'training') % Train a mapping.
[a,classf,nfeat,nclassf] = deal(argin{:});
isvaldset(a,1,1);
[m,k] = size(a);
if isempty(nfeat)
nfeat = max(round(m/10),2); % use at least 2D feature spaces
end
if isempty(nclassf)
nclassf = max(ceil(k/nfeat),10); % use at least 10 classifiers
end
if nfeat >= k % allow for small feature sizes (k < nfeat)
nfeat = k; % use all features
nclassf = 1; % compute a single classifier
end
nsets = ceil(nfeat*nclassf/k);
featset = zeros(k,nsets);
for j=1:nsets
featset(:,j) = randperm(k)';
end
featset = featset(1:nfeat*nclassf);
featset = reshape(featset,nclassf,nfeat);
w = [];
s = sprintf('Compute %i classifiers: ',nclassf);
prwaitbar(nclassf,s);
for j=1:nclassf
prwaitbar(nclassf,j,[s num2str(j)]);
w = [w; a(:,featset(j,:))*classf];
end
prwaitbar(0)
w = prmapping(mfilename,'trained',{w,featset},getlablist(a),k,getsize(a,3));
else % Evaluation
[a,w] = deal(argin{1:2});
wdata = getdata(w);
w = wdata{1};
featset = wdata{2}';
w = a(:,featset(:))*w;
end
return