mb_knnsummary.m

 function[summary]=mb_knnsummary(knntrials)
%  MB_KNNSUMMARY - Summarize kNN performance over several trials
%
%  [summary] = MB_KNNSUMMARY(KNNTRIALS)
%
%    Outputs:
%     SUMMARY.CMAT - confusion matrix derived (% correct form) from 
%             all of the elements in knntrials (number correct form)
%     SUMMARY.CMAT_NOUNK - confusion matrix for classification attempts
%     SUMMARY.CRATE - Two elements: 
%              1. average of the diagonal elements in CMAT
%              2. average of the diagonal elements in a confusion matrix
%              generated by excluding the unknown category (i.e. what 
%              fraction of the classification attempts were correct.
%
%    Inputs:
%     KNNTRIALS - cell array output from mb_knntrainpicktest
%
%
% M. Boland - 15 Apr 1999


% $Id: mb_knnsummary.m_tmp,v 1.1 1999/06/26 14:16:36 boland Exp $

if(~iscell(knntrials))
  error('KNNTRIALS must be a cell array containing output 
	from mb_knntrainpicktest') ;
end

%
% Sum all confusion matrices in knntrials
confusion = zeros(size(knntrials{1}.cmat)) ;
crates=[] ;

for i=1:length(knntrials)
  confusion = confusion + knntrials{i}.cmat ;
  crates = [crates ; knntrials{i}.crate] ;
end

%
% Calculate the confusion matrix as percentages 
%  rather than numbers of samples
summary.cmat = confusion ./ (sum(confusion')' * ones(1,size(confusion,2))) ;

%
% Same calculations for the matrix without the unknown classes 
%  (i.e. what fraction of the non-unknown samples are correctly classified)
confusion_nounk = confusion(:,1:(end-1)) ;
summary.cmat_nounk = confusion_nounk ./ (sum(confusion_nounk')' * ...
                   ones(1,size(confusion_nounk,2))) ;

numunknown = sum(confusion(:,end)) ;
numcorrect = sum(diag(confusion_nounk)) ;

summary.Pc_mean=[mean(crates)] ;
summary.Pc_var =[var(crates)] ;