%ADABOOSTC
%
% [W,V,ALF] = ADABOOSTC(A,CLASSF,N,RULE,VERBOSE);
%
% INPUT
% A Dataset
% CLASSF Untrained weak classifier
% N Number of classifiers to be trained
% RULE Combining rule (default: weighted voting)
% VERBOSE Suppress progress report if 0 (default)
%
% OUTPUT
% W Combined trained classifier
% V Cell array of all classifiers
% Use VC = stacked(V) for combining
% ALF Weights
%
% DESCRIPTION
%
% Computation of a combined classifier according to adaboost.
%
% In total N weighted versions of the training set A are generated
% iteratevely and used for the training of the specified classifier.
% Weights, to be used for the probabilities of the objects in the training
% set to be selected, are updated according to the Adaboost rule.
%
% The entire set of generated classifiers is given in V.
% The set of classifier weigths, according to Adaboost is returned in ALF
%
% Various aggregating possibilities can be given in
% the final parameter rule:
% []: WVOTEC, weighted voting.
% VOTEC voting
% MEANC sum rule
% AVERAGEC averaging of coeffients (for linear combiners)
% PRODC product rule
% MAXC maximum rule
% MINC minimum rule
% MEDIANC median rule
%
% REFERENCE
% Ji Zhu, Saharon Rosset, Hui Zhou and Trevor Hastie,
% Multiclass Adaboost. A multiclass generalization of the Adaboost
% algorithm, based on a generalization of the exponential loss.
% http://www-stat.stanford.edu/~hastie/Papers/samme.pdf
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% MAPPINGS, DATASETS
% 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
% (Multiclass correction by Marcin Budka, Bournemouth Univ., UK)
%function [W,V,alf] = adaboostc(a,clasf,n,rule,verbose)
function [out,V,alf] = adaboostc(varargin)
%% INITIALISATION
argin = setdefaults(varargin,[],nmc,100,[],0);
if mapping_task(argin,'definition')
out = define_mapping(argin,'untrained','Adaboost');
%% TRAINING
elseif mapping_task(argin,'training')
[a,clasf,n,rule,verbose] = deal(argin{:});
[m,k,c] = getsize(a);
V = [];
laba = getlab(a);
u = ones(m,1)/m; % initialise object weights
alf = zeros(1,n); % space for classifier weights
isseparable = 0; % check if we can make 0 error
if verbose && k == 2
figure(verbose);
scatterd(a);
end
%% generate n classifiers
for i = 1:n
b = gendatw(a,u,m); % sample training set
b = setprior(b,getprior(a)); % use original priors
w = b*clasf; % train weak classifier
ra = a*w; % test weak classifier
if verbose && k == 2
plotc(w,1); drawnow
end
labc = labeld(ra);
diff = sum(labc~=laba,2)~=0; % objects erroneously classified
erra = sum((diff).*u); % weighted error on original dataset
if (erra==0)
isseparable = 1;
V = w;
break;
end
if (erra < (1-1/c)) % if classifier better then random guessing...
alf(i) = 0.5*(log((1-erra)/erra) + log(c-1));
correct = find(diff==0); % find correctly classified objects
wrong = find(diff==1); % find incorrectly classified objects
u(correct) = u(correct)*exp(-alf(i)); % give them the ...
u(wrong) = u(wrong)*exp(alf(i)); % proper weights
u = u./sum(u); % normalize weights
else
alf(i) = 0;
end
if verbose
disp([erra alf(i) sum(alf)])
end
V = [V w]; % store all classifiers
end
%% combine and return
if isseparable
W = V;
W = setname(W,['Boosted ',getname(V)]);
else
if isempty(rule)
W = wvotec(V,alf); % default is weighted combiner
else
W = traincc(a,V,rule); % otherwise, use user supplied combiner
end
end
if verbose > 0 && k == 2
plotc(W,'r',3)
ee = a*W*testc;
title(['Error: ', num2str(ee)]);
end
out = W;
else
error('Illegal call')
end
return