add

add demo for kernel kmeans

chapter06/knKmeans.m

@@ -1,7 +1,7 @@
-function [label, model, energy] = knKmeans(X, init, kn)
+function [label, model, energy] = knKmeans(K, init)
 % Perform kernel kmeans clustering.
 % Input:
-%   K: n x n kernel matrix
+%   K: n x n data matrix
 %   init: either number of clusters (k) or initial label (1xn)
 % Output:
 %   label: 1 x n sample labels
@@ -10,17 +10,13 @@
 % Reference: Kernel Methods for Pattern Analysis
 % by John Shawe-Taylor, Nello Cristianini
 % Written by Mo Chen (sth4nth@gmail.com).
-n = size(X,2);
+n = size(K,2);
 if numel(init)==1
     k = init;
     label = ceil(k*rand(1,n));
 elseif numel(init)==n
     label = init;
 end
-if nargin < 3
-    kn = @knGauss;
-end
-K = kn(X,X);
 last = zeros(1,n);
 while any(label ~= last)
     [~,~,last(:)] = unique(label);   % remove empty clusters

demo/ch06/knKmeans_demo.m

@@ -0,0 +1,23 @@
+%% Kernel kmeans with linear kernel is equivalent to kmeans
+close all; clear;
+d = 2;
+k = 3;
+n = 200;
+[X, y] = kmeansRnd(d,k,n);
+init = ceil(k*rand(1,n));
+K = knLin(X,X);
+label = knKmeans(K,init);
+
+label0 = kmeans(X,init);
+maxdiff(label,label0)
+plotClass(X,label);
+%% Kernel kmeans with Gaussian Kernel for nonlinear data
+x1 = linspace(0,pi,n/2);
+x2 = sin(x1);
+X = [x1,x1+pi/2;
+    x2,-x2];
+
+K = knGauss(X,X,0.4);
+label = knKmeans(K,2);
+figure;
+plotClass(X,label);