%LABELD Find labels of classification dataset (perform crisp classification)
%
% LABELS = LABELD(Z)
% LABELS = Z*LABELD
% LABELS = LABELD(A,W)
% LABELS = A*W*LABELD
% LABELS = LABELD(Z,THRESH)
% LABELS = Z*LABELD([],THRESH)
% LABELS = LABELD(A,W,THRESH)
% LABELS = A*W*LABELD([],THRESH)
%
% INPUT
% Z Classification dataset, or
% A,W Dataset and classifier mapping
% THRESH Rejection threshold
%
% OUTPUT
% LABELS List of labels
%
% DESCRIPTION
% Returns the labels of the classification dataset Z (typically the result
% of a mapping or classification A*W). For each object in Z (i.e. each row)
% the feature label or class label (i.e. the column label) of the maximum
% column value is returned.
%
% Effectively, this performs the classification. It can also be considered
% as a conversion from soft labels Z to crisp labels.
%
% When the parameter THRESH is supplied, then all objects which
% classifier output falls below this value are rejected. The returned
% label is then NaN or a string with spaces (depending if the labels are
% numeric or string). Because the output of the classifier is used, it
% is recommended to convert the output to a posterior prob. output using
% CLASSC. (David Tax, 27-12-2004)
%
% SEE ALSO
% MAPPINGS, DATASETS, TESTC, PLOTC
% 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 labels = labeld(a,w,thresh)
% Add the possibility to reject objects for which the posterior is
% too low:
if (nargin < 3)
thresh = [];
end
if (nargin == 2) & isa(w,'double') % we did something like labeld(z,0.3)
thresh = w;
w = [];
end
if (nargin < 2)
w = [];
end
if (nargin == 0) | isempty(a)
% Untrained mapping.
labels = prmapping(mfilename,'fixed',{thresh});
elseif isempty(w)
if (isdatafile(a)) % datafile needs to process objects separately
labels = cell(size(a,1),1);
next = 1;
while next > 0
[b,next,J] = readdatafile(a,next);
labs = feval(mfilename,b,[],thresh);
for i=1:length(J)
labels{J(i)} = labs(i,:);
end
end
if isstr(labels{1})
labels = char(labels);
else
labels = cell2mat(labels);
end
return
end
% In a classified dataset, the feature labels contain the output
% of the classifier.
[m,k] = size(a); featlist = getfeatlab(a);
Jrej = []; % as a start, we don't reject objects
if (k == 1)
% If there is one output, assume it's a 2-class discriminant:
% decision boundary = 0.
J = 2 - (double(a) >= 0);
if ~isempty(thresh)
warning('Inproper thresholding of the 2-class dataset, please use classc.');
end
else
% Otherwise, pick the column containing the maximum output.
[dummy,J] = max(+a,[],2);
% Reject the objects which have posteriors lower than the
% threshold
if ~isempty(thresh)
Jrej = find(dummy<thresh);
end
end
labels = featlist(J,:);
% Take care for the rejected objects:
if ~isempty(Jrej)
if isa(featlist,'double')
labels(Jrej) = NaN;
elseif isa(featlist,'char')
% labels(Jrej,:) = repmat(' ',size(featlist(1,:)) ); % Trick!:-)
labels(Jrej,:) = repmat(' ', length(Jrej), size(labels,2) ); % Trick!:-)
else
error('The featlist of A is confusing. Cannot make a reject label');
end
end
else
% Just construct classified dataset and call again.
labels = feval(mfilename,a*w,thresh);
end
return