weightedknnrule_vc.m

最新的模式识别分类工具箱,希望对朋友们有用!

% Learns classifier and classifies test set
% using weighted k-NN rule
% Usage
%      [trainError, testError, estTrainLabels, estTestLabels] = ...
%           weightedKNNRule_VC(trainFeatures, trainLabels,params ,testFeatures, testLabels)
% where
%
% Inputs:
% 	trainFeatures  - the training set vectors, one vector per column
%	trainLabels    - the labels of the above
%       params         - <k>, the number of nearest neighbors
%                        <weightFile>, a file with the weights, a
%                           a vector having k (non-increasing) entries. 
%                           Entry 1 is the weight of the nearest neighbor,
%                           while entry k is the weight of the farthest neighbor
%                         IF weightFile IS NOT PROVIDED, DEFAULT weights are
%                           provided using the MacLeod, Luk and Titterington
%                           weights with parameters s = 2k, and alpha = .5
%                           (see lecture notes)
%                         to avoid providing <weightfile> there are 2 alternatives
%                          1. in the user interface, use '' to indicate no
%                              weight file
%                          2. provide a file name of a non-existing file
%                           (e.g.,'foononexist')
%                         An example of weight file, "exampleWeightFile"
%                           is provided as a reference
%       testFeatures   - test set, one column per vector
%       testLabels     - labels  for test set%
% Outputs
%	trainError     - the error rate on the training set (one entry per
%	                   class + total error)
%	testError      - the error rate on the test set (one entry per class
%	                  + total error)
%       estTrainLabels - the labels produced by the algorithm for the
%                          training samples
%       estTestLabels - the labels produced by the algorithm for the
%                          test samples


function  [trainError, testError, estTrainLabels, estTestLabels] = ...
    weightedKNNRule_VC(trainFeatures, trainLabels,algParam ,testFeatures, testLabels)

%
%
[Nclasses, classes]  = find_classes([trainLabels(:);testLabels(:)]); % Number of classes in labels
[Dim, Nsam]          = size(trainFeatures);

comma_loc  = findstr(algParam,',');
Knn        = str2num(algParam(2:comma_loc(1)-1));
weightfile = algParam((comma_loc(1)+2):(length(algParam)-2));


%--------------------------------
% opening the weights file
if ( weightfile ~= [])
  fp = fopen(weightfile,'r');

  if ( fp ~= -1)
    weights = fscanf(fp,'%f');
    fclose(fp);
    haveWeights =1;
    if ( length(weights) > Knn)
      weights = weights(1:Knn);
      weights = weights(:); % column vector
    elseif (length(weights) < Knn)
      weights = [weights(:);zeros(Knn-length(weights),1)]; %pad with zeros
    else
      weights = weights(:); % column vector
    end
  else
    haveWeights = 0;
  end
else
  haveWeights =0;
end


hm = findobj('Tag', 'Messages'); 
fprintf('Weighted %d-N-N Rule: Classifying Training Set\n',Knn);
if (isempty(hm)==0)
  s = sprintf('Weighted %d-N-N Rule: Classifying Training Set\n',Knn);
  set(hm,'String',s);
  refresh;
end

% --------------------------------
% classifies training set

[Dim,Nsamples] = size(trainFeatures);
Onen   = ones(1,Nsamples);

for sam=1:Nsamples
  aux    = trainFeatures(:,sam) *Onen;
  sdiff  = trainFeatures - aux;
  sdiff  = sdiff .* sdiff;

  % computes the squared distances, no need to take square rots
  distsq = sum(sdiff);
  [sorted_dist, indices] = sort(distsq);

  if length(trainLabels) <= Knn
   k_nearest = trainLabels;
 else
   k_nearest = trainLabels(indices(1:Knn));
 end
 
 % If necessary, computes the  weight vector
 if haveWeights == 0,
   k1 =2*Knn;
   % a small scrambling to deal with k > n/2 !
   if ( k1 < length(trainLabels))
     dmax = sqrt ( 2 * max(sorted_dist));
     if ( k1 < Knn) 
       k1 = length(trainLabels);
     end
   else
     dmax = sqrt(sorted_dist(k1));
     k1   = Knn;
   end
   alpha   = 0.5;
   auxDist = sqrt(sorted_dist(1:Knn));
   weights = ((dmax - auxDist) + alpha * (dmax - auxDist(1))) ./ ... 
       ((1+alpha)*(dmax - auxDist(1)));
 end

 counts = zeros(1,Nclasses);
 for i=1:Knn,
   label = k_nearest(i);
   label = find(classes == label);
   counts (label) =   counts (label) + weights(i);
 end

 [foo,target] =  max(counts);
 estTrainLabels(sam) = classes(target);
end

% ================================
% classifies the test set 

hm = findobj('Tag', 'Messages'); 
fprintf('Weighted %d-N-N Rule: Classifying Test Set\n',Knn);
if (isempty(hm)==0)
  s = sprintf('Weighted %d-N-N Rule: Classifying Test Set\n',Knn);
  set(hm,'String',s);
  refresh;
end

[Dim,Nsamples] = size(testFeatures);

for sam=1:Nsamples
  aux    = testFeatures(:,sam) *Onen;
  sdiff  = trainFeatures - aux;
  sdiff  = sdiff .* sdiff;

  % computes the squared distances, no need to take square rots
  distsq = sum(sdiff);
  [sorted_dist, indices] = sort(distsq);

  if length(trainLabels) <= Knn
   k_nearest = trainLabels;
 else
   k_nearest = trainLabels(indices(1:Knn));
 end
 
 % If necessary, computes the  weight vector
 if haveWeights == 0,
   k1 =2*Knn;
   % a small scrambling to deal with k > n/2 !
   if ( k1 < length(trainLabels))
     dmax = sqrt ( 2 * max(sorted_dist));
     if ( k1 < Knn) 
       k1 = length(trainLabels);
     end
   else
     dmax = sqrt(sorted_dist(k1));
     k1   = Knn;
   end
   alpha   = 0.5;
   auxDist = sqrt(sorted_dist(1:Knn));
   weights = ((dmax - auxDist) + alpha * (dmax - auxDist(1))) ./ ... 
       ((1+alpha)*(dmax - auxDist(1)));
 end
 
 counts = zeros(1,Nclasses);
 for i=1:Knn,
   label = k_nearest(i);
   label = find(classes == label);
   counts (label) =   counts (label) + weights(i);
 end

 [foo,target] =  max(counts);
 estTestLabels(sam) = classes(target);
end
%
%

trainError = computeError( classes, trainLabels, estTrainLabels);
testError  = computeError( classes, testLabels, estTestLabels);