knn_rule.m

最新的模式识别分类工具箱,希望对朋友们有用!

% Classifies input using k-NN rule
% Usage
%      label = Knn_Rule(inputSample, TrainingSamples, TrainingLabels, k);
% where
%    inputSample     is a ROW vector (1xd) 
%                      containing the sample to be classified
%    TrainingSamples is a matrix (nxd) containing 1 training vector
%                      with d features in each of its n rows, where
%                      n is the training set size
%    TrainingLabels  is a COLUMN vector (nx1) containing the labels of
%                      the training samples
%    k               is the number of nearest neighbors.
%
%  NOTE: the above conventions on rows vs. columns are not extremely
%  important, since the routine will automatically adjust to cases
%  where they are violated.
%  HOWEVER: if the number of vectors in the training set is the same as
%  the number of dimensions, the algorithm will not be able to tell which
%  is which!
%
% A caveat:
%   if at all possible, use the set (1,...,C) as class labels.
%   this implementation will not work that well if you use very large 
%   or very small (negative) numbers as class labels (like 10^23, for example)
%
%
%  Used by Store_Grabbag
%


function target = Knn_Rule(Sample, trainingVectors, classLabels, Knn)

% Date     Name               Change
% 03/18/02 Vittorio Castelli  Now works for any number of features,
%                             any number of classes. 
%                             Added documentation


[a,b] = size(trainingVectors);
if ( a ~= length(classLabels) )
  if ( b ~= length(classLabels))
    fprintf('Apparent inconsistency between arguments 2 and 3\n');
    fprintf('Argument 2 should contain the training set, with vectors as rows\n');
    fprintf('Argument 3 should contain the labels and have length equal to the number of rows of argument 2\n');
    target = -1;
    return;
  else
    % Rearranges the training vectors according to the
    % desired order
    trainingVectors = trainingVectors';
    [a,b] = size(trainingVectors);
  end
end

% computes the squared differences between the entries
% of each training sample and the corresponding entries
% of the sample to be classified

[c d] = size(Sample);
if (c > d)
  % rearranges the sample as a row vector
  Sample = Sample';
  [c d] = size(Sample);
end
if (c > 1)
  fprintf('This algorithm only classifies one vector at a time\n');
  target = -1;
  return;
end


aux    = ones(a,1) * Sample;
sdiff  = trainingVectors - aux;
sdiff  = sdiff .* sdiff;

% computes the squared distances, no need to take square rots
distsq = sum(sdiff,2);

[sorted_dist, indices] = sort(distsq);
z = min(classLabels);
if (z <= 0)
  classLabels = classLabels - z+1;
end
Z      = max(classLabels);
counts = zeros(1,Z);


if length(classLabels) <= Knn
   k_nearest = classLabels;
else
   k_nearest = classLabels(indices(1:Knn));
end



for i=1:Knn,
  counts (k_nearest(i)) =   counts (k_nearest(i))+1;
end
[foo,target] =  max(counts);

if (z<=0)
  target = target-1+z;
end