This function performs kernel kmeans algorithm. When the linear kernel (i.e., inner product) is used, the algorithm is equivalent to standard kmeans algorithm. Several nonlinear kernel functions are also provided. Upon request, I also include a prediction function for out-of-sample inference. Please try following code for a demo:
clear; close all;
d = 2;
k = 3;
n = 500;
[X,label] = kmeansRnd(d,k,n);
init = ceil(k*rand(1,n));
[y,mse,model] = knKmeans(X,init,@knLin);
plotClass(X,y)
idx = 1:2:n;
Xt = X(:,idx);
t = knKmeansPred(model, Xt);
plotClass(Xt,t)
This function is now a part of the PRML toolbox (http://www.mathworks.com/matlabcentral/fileexchange/55826-pattern-recognition-and-machine-learning-toolbox).
Jung, Tajana, I found the same problem in my Matlab version R2014b.
The function "unique" on line 18 changes the variable "label" from a row vector to a column vector.
This caused the error in my case.
To fix it, transpose the variable "label" after line 18.
Jung and Tajana, I ran the sample code without any problems but I'm wondering if your data array is the right size (2 X 1000).
In either case, you might want to see if the 'any' function in your MATLAB version can compare a vector to a scalar or to another vector. If you get the same 'Matrix dimensions must agree' error then that may be the problem.
Hello I encountered the same problem as john luckily i had the book, I added the following code. The energy is the sum squared clustering cost function. I have been optimizing my kernel hyper-parameters to minimise this energy. Been working fairly well so thanks. Not an expert so could be wrong.
A=zeros(size(S'));
for i=1:1:size(K,1)
A(i,label(i))=1;
end
D=diag(1./sum(A));
energy = trace(K)-trace(sqrt(D)*A'*K*A*sqrt(D));
Hi mathieu,
As indicated in the description, this algorithm is explained in
reference: [1] Kernel Methods for Pattern Analysis
by John Shawe-Taylor, Nello Cristianini
Mathieu, you can refer to machine learning and pattern recognition by Bishop, 2005. Alternatively this is for free: www-stat.stanford.edu/~hastie/Papers/ESLII.pdf
I see, reading the code I do not manage to understand what are the principles behind the algorithm. Do you have a reference that I could get from the web or do you advise to buy the book ?
This happens for standard kmeans too, which is caused by the nature of the algorithm. The reason is that when you set a very big number for k, after several iterations, some clusters might become empty.
It seems that if I request N clusters, the algorithms outputs k clusters, k<=N clusters and most of the time k<<N. I was wondering if this is by construction. If yes, could provide me with an explanation ?
Updates
25 Dec 2009
1.1
add sample data and detail description
30 Sep 2010
1.2
remove empty clusters
03 Feb 2012
1.5
fix a minor bug of returning energy
03 Feb 2012
1.6
n/a
03 Feb 2012
1.7
Improve the code and fix a bug of returning energy
07 Mar 2016
1.7
fix incompatibility issue due the stupid API change of function unique()