This is a fully vectorized version kmedoids clustering methods (http://en.wikipedia.org/wiki/K-medoids). It is usually more robust than kmeans algorithm. Please try following code for a demo:
close all; clear;
d = 2;
k = 3;
n = 500;
[X,label] = kmeansRnd(d,k,n);
y = kmedoids(X,k);
Input data are assumed COLUMN vectors!
You can only visualize 2d data!
This function is now a part of the PRML toolbox (http://www.mathworks.com/matlabcentral/fileexchange/55826-pattern-recognition-and-machine-learning-toolbox)
Thanks for this code, but for some datasets it's hypersensitive to rounding errors: occasionally the slightly nonzero entries along D's diagonal lead to results that are surprisingly far from correct (eg the chance that given medoids are chosen increases or decreases). I can email code to demonstrate if needed.
The problem is fixed by adding a for-loop after D is defined:
D(i,i) = 0;
To confirm that this yields 'correct' results, I perturbed one of the problematic datasets by a tiny amount and compared medoids before-and-after.
Hope that helps.
Even though the code is lightning fast, the solution is not the proper one, hence this code is useless. See http://i.imgur.com/VNY73l7.png for example output of k = 6 for a naive (and very slow) implementation of the algorithm, and this submission. Obviously the naive is correct, this submission is incorrect.
Nonetheless thanks for the effort! It would be great if you could produce correct code that is still as fast.
I tested calling to kmedoids() function, it always returns
label = 2 1
energy = 0
index = 2 1
I believe it does not work. Thanks anyway.
Nice code. Very helpful. However, if the number of input data is large, then the D matrix is too large (n^2) which may make memory overflow.
Simple and elegant code, thanks!
If I want to use the cosine distance between two vectors, What shold I do?
if i want to apply kmedoids agorithm for X data in function [label, energy, index] = kmedoids(X,k) 3599*11 size data then it is not working properly can anybody of you give idea for this?
Very compact and efficient coding. Nice job!
However, there is an error when k=1. I suggest to add the third parameter (dimension) to the calls of function min:
(Line 8): [~, label] = min(D(randsample(n,k),:),,1);
(Line 11): [~, label] = [~, index] = min(D*sparse(1:n,label,1,n,k,n),,1);
Not working for my data as well.
My data was of the dimension 17-by-71.
wanted to find 4 or 6 clusters.
and spread did not work.
Undefined function 'randsample' for input arguments of type 'double'.
Error in kmedoids (line 8)
[~, label] = min(D(randsample(n,k),:));
better post your error message
Dear Sir Pls help!!! My Matlab version is Matlab7.6.0..Is this the reason?
Pls reply when you free!!!
SIMPLY NOT working.
improved numerical stability, remove empty clusters
fix bug for k=1
significantly simplify the code