gkclust.m

这是一个聚类的工具箱,可以用到数据挖掘等高级应用中去.

function result = GKclust(data,param)

%checking the parameters given
f0=param.c;
X=data.X;

[N,n] = size(X);
[Nf0,nf0] = size(f0);
X1 = ones(N,1);
%default parameters
if exist('param.m')==1, m = param.m;else m = 2;end;
if exist('param.e')==1, e = param.m;else e = 1e-4;end;
if exist('param.ro')==1, rho=param.ro;
else 
    if max(Nf0,nf0) == 1
        rho = ones(1,param.c);
    else
        rho = ones(1,size(f0,2));
    end
end
if exist('param.gamma')==1, gamma = param.gamma;else gamma = 0;end;
if exist('param.beta')==1, beta = param.beta;else beta = 1e15;end;



% Initialize fuzzy partition matrix
rand('state',0)
if max(Nf0,nf0) == 1, 		% only number of cluster given
  c = f0;
  mm = mean(X);
  aa = max(abs(X - ones(N,1)*mm));
  v = 2*(ones(c,1)*aa).*(rand(c,n)-0.5) + ones(c,1)*mm; 
  for j = 1 : c,
    xv = X - X1*v(j,:);
    d(:,j) = sum((xv.^2),2);
  end;
  d = (d+1e-10).^(-1/(m-1));
  f0 = (d ./ (sum(d,2)*ones(1,c)));
else        %if the partition matrix was given
  c = size(f0,2);
  fm = f0.^m; sumf = sum(fm);
  v = (fm'*X)./(sumf'*ones(1,n));
end;

f = zeros(N,c);                 % partition matrix
iter = 0;                       % iteration counter
A0= eye(n)*det(cov(X)).^(1/n);  % "identity" matr.



% Iterate
while  max(max(f0-f)) > e
  iter = iter + 1;
  f = f0;
  % Calculate centers
  fm = f.^m; sumf = sum(fm);
  v = (fm'*X)./(sumf'*ones(1,n));
  for j = 1 : c,
    xv = X - X1*v(j,:);
    % Calculate covariance matrix for each clusters
    A = ones(n,1)*fm(:,j)'.*xv'*xv/sumf(j);
    %Covariance estimating for the GK algorithm
    A =(1-gamma)*A+gamma*(A0.^(1/n));
    if cond(A)> beta;
        [ev,ed]=eig(A); edmax = max(diag(ed));
        ed(beta*ed < edmax) = edmax/beta;
        A = ev*diag(diag(ed))*inv(ev);
    end;
    %Calculate distances
    M = (1/det(pinv(A))/rho(j))^(1/n)*pinv(A);
    %M(:,:,j) = (det(A)/rho(j)).^(1/n)*pinv(A);
    d(:,j) = sum((xv*M.*xv),2);
  end
    
  distout=sqrt(d);
  
  J(iter) = sum(sum(f0.*d));           %calculate objective function
  % Update f0
  d = (d+1e-10).^(-1/(m-1));
  f0 = (d ./ (sum(d,2)*ones(1,c)));
   
end             %end of iteration period

fm = f.^m; sumf = sum(fm);

P = zeros(n,n,c);             % covariance matrix
M = P;                          % norm-inducing matrix
V = zeros(c,n);                % eigenvectors
D = V;                          % eigenvalues

% calculate P,V,D,M
for j = 1 : c,                        
    xv = X - ones(N,1)*v(j,:);
    % Calculate covariance matrix
    A = ones(n,1)*fm(:,j)'.*xv'*xv/sumf(j);
    % Calculate eigen values and eigen vectors
    [ev,ed] = eig(A); ed = diag(ed)';
    ev = ev(:,ed == min(ed));
    % Put cluster info in one matrix
	P(:,:,j) = A;
    M(:,:,j) = (det(A)/rho(j)).^(1/n)*pinv(A);
    V(j,:) = ev';
    D(j,:) = ed;
end



%result outputs
result.data.f = f0;
result.data.d = distout;
result.cluster.v = v;
result.cluster.P = P;
result.cluster.M = M;
result.cluster.V = V;
result.cluster.D = D;
result.iter = iter;
result.cost = J;