K-means: see the initial centroids

Hi everyone,

Is it possible to see the randomly selected initial centroids when using the built-in kmeans function? I want to keep track of the centroids in each iteration.

Thanks.

Answer 1

[Massimo Zanetti]

Yes it is possible.

kmeans allows you to set option parameters via the statset function. In the help page kmeans there some examples on how using stateset.

Set the parameter 'OutputFcn' as shown in this tutorial, for displaying iteration results in iterative solvers: https://it.mathworks.com/help/matlab/math/output-functions.html

[Phu Lai]

It seems that Output Function only works with optimization functions such as fminbnd, fminsearch or fzero. I tried implementing 'OutputFcn' like this but not working:

function [history] = kmeansIter(data,k)
    history = [];
    options = statset('OutputFcn', @myoutput,'MaxIter', 500, 'Display', 'final');
    [clustIDX, clusters, interClustSum, Dist] = kmeans(data, k, 'options',options, ...
            'distance','sqEuclidean', 'EmptyAction','singleton', 'replicates',3);
    function stop = myoutput(x,optimvalues,state);
        stop = false;
        if isequal(state,'iter')
          history = [history; x];
        end
    end
end

Can you please have a look to see if I've done anything wrong? Thanks very much.

[Massimo Zanetti]

Ouch. Surprisingly, the stateset function has only some options available with kmeans. This is quite disappointing.

Therefore it seems there is no other way to circumvent your problem, than performing separate runs of kmeans (all with maxiter parameter set to 1) and save the results. Here is an example:

load fisheriris
X = meas(:,3:4);

MaxIt=4;
C = cell(1,MaxIt);
I = cell(1,MaxIt);

rng(1); % For reproducibility
[I{1},C{1}] = kmeans(X,3,'Display','iter','MaxIter',1);
for k=2:MaxIt
    [I{k},C{k}] = kmeans(X,3,'Display','iter','Start',C{k-1},'MaxIter',1);
end
fig=figure;
for k=1:4
    figure(fig);
    gscatter(X(:,1),X(:,2),I{k}); grid on; hold on;
    plot(C{k}(:,1),C{k}(:,2),'xk','MarkerSize',12,'LineWidth',5); hold off;
    pause(.5);
end

[Phu Lai]

Thank you. I ended up following that approach. Also, do you know how to see the first initial centroids? (Let's say, C{0} in your example)

[Massimo Zanetti]

Nope.. I have tried setting 'MaxIter'=0, but it seems it is the same as setting 'MaxIter'=1. The kmeans function is a black box :(