kmeans_dd.m

模式识别工具箱,希望对大家有用!

%KMEANS_DD k-means data description.
% 
%       W = kmeans_dd(A,fracrej,K)
% 
% Train a k-means method with K prototypes on dataset A. Parameter
% fracrej gives the fraction of the target set which will be rejected.
% 
% This method requires the k-means from the NetLab toolbox. Here the
% means are optimized using a EM algorithm. To avoid confusion with
% the PRTools kmeans, the NetLab kmeans is copied to mykmeans and this
% version is used.
% 
% See also datasets, mappings, dd_roc

% Copyright: D. Tax, R.P.W. Duin, [e-mail]
% Faculty of Applied Physics, Delft University of Technology
% P.O. Box 5046, 2600 GA Delft, The Netherlands
  
function [W,out] = kmeans_dd(a,fracrej,K,epochs)

if nargin < 4,  epochs = 1000; end
if nargin < 3 | isempty(K), K = 5; end
if nargin < 2 | isempty(fracrej), fracrej = 0.05; end
if nargin < 1 | isempty(a) % empty dd
  W = mapping('kmeans_dd',{fracrej,N});
  return
end

if isa(fracrej,'double')           %training

  if ~isa(a,'dataset')              %train on training set
    error('KMEANS_DD: I need a dataset to train');
  end
  a = oc_set(a);     % make sure a is an OC dataset
  [nlab,lablist,m,k,c] = dataset(a);

  % train it:
  options = 0; options(14) = epochs;
  w0 = rand(K,k);
  w = mykmeans(w0, +a, options);
  w = setdiff(w,w0,'rows');
  w = unique(w,'rows');
%  if max(size(w))<2
%    error('Did not find a solution');
%  end

  % obtain the threshold:
  d = min(dist(+a,w'),[],2);
  if (size(d,2)~=1)
    d = d';
  end
  thr = -threshold(-d,fracrej);

  %and save all useful data:
  W.w = w;
  W.threshold = thr;
  W.scale = mean(d);
  W = mapping('kmeans_dd',W,str2mat('target','outlier'),k,c);

else                               %testing

  [W,classlist,type,k,c] = mapping(fracrej);  % unpack
  [nlab,lablist,m,k,c,p] = dataset(a);

  %compute:
  out = min(dist(W.w,a'))';
  if max(size(out))==1
    out
  end
  newout = dist2dens([out ones(m,1)*W.threshold],W.scale);

  W = dataset(newout,getlab(a),classlist,p,lablist);
end

return