lvq_dd.m

模式识别工具箱,希望对大家有用!

% w = lvq_dd(x,fracrej,N,learningrate,epochs);
% 
% try to detect outliers using LVQ
% when labx is given, the lvq tries more or less to avoid the objects
% with labx == -1
%
% Problem is, that in the new Matlab6 neural network toolbox, the
% performance on the data is checked first. And if all data is from
% one class, we immediately classify everything correctly, and nothing
% is optimized further.
function W = lvq_dd(x,fracrej,N,learningrate,epochs);
if (nargin<5)
  epochs=1000;
end
if (nargin<4)
  learningrate=0.05;
end
if (nargin<3)
  N=5;
end

if ((nargin==2) & (size(x,2)==size(N,2)))  %testing fase
  w = x;
  Dx = min(dist(N,w'),[],2);
  return;
end

[nrx,dim] = size(x);
pc = [length(find(labx==1)) length(find(labx==-1))]/length(labx);
minmax = [min(x)' max(x)'];
if (pc(2)==0)
  pc = 1.0;
  labx = ones(1,size(x,1));
else
  labx = ind2vec((3*ones(size(labx)) - labx)/2);
end
net = newlvq(minmax,N,pc);
net.trainParam.epochs = epochs;
net.trainParam.lr = learningrate;
net.trainParam.show = inf;

net = train(net,x',labx);
w = net.IW{1,1};
clI = vec2ind(net.LW{2,1});
dw = (w - ones(size(w,1),1)*mean(minmax')).^2;
I = find(sum(dw,2)<0.001);
w(I,:) = [];
clI(I) = [];

w = w(clI==1,:);
Dx = min(dist(x,w'),[],2);

return