running k-means and getting different results run after run?

I am running k-means clustering algorithm on a data, and I don't understand why I am getting different silhouette plots each time I run this. Is there a way to stabilise this? (or set the number of iterations) so I get the same results?

[Pieter Hamming]

Generally, k-means clustering generates different results depending on starting conditions; that's why you'd normally run it a few times with randomized starting conditions.

I don't know the specific MATLAB implementation of k-means, but I'd assume it runs multiple times with random initialization.

Could you give an example of your data/results so we can see if the issue is in the (application of the) algorithm or in the data?

[cgo]

<<

These are two results of the the same data, and the same number of clusters (2). Is the data just that bad? Or I am not getting something right here?

Thanks for your insights.

>>

[cgo]

Answer 1

[Image Analyst]

That's normal. Specify 'Replicates' to get convergence.

% Do kmeans clustering on the gray scale image.
grayLevels = double(grayImage(:)); % Convert to column vector.
[clusterIndexes, clusterCenters] = kmeans(grayLevels, numberOfClusters,...
  'distance', 'sqEuclidean', ...
  'Replicates', 2);
labeledImage = reshape(clusterIndexes, rows, columns);

See attached demo.

[Mehmet Volkan Ozdogan]

Hi,

I have the same problem

my code is;

x=xlsread('ucd1.xlsx', 'A:A');

y=xlsread('ucd1.xlsx', 'B:B');

z=xlsread('ucd1.xlsx', 'C:C');

a=xlsread('ucd2.xlsx', 'A:A');

b=xlsread('ucd2.xlsx', 'B:B');

c=xlsread('ucd2.xlsx', 'C:C');

xyz=[x y z];

abc=[a b c];

[idx1,C1] = kmeans(xyz,150,'Replicates',3);

[idx2,C2] = kmeans(abc,150,'Replicates',3);

plot3(C1(:,1),C1(:,2),C1(:,3),'kx','color','b')

hold on

plot3(C2(:,1),C2(:,2),C2(:,3),'ko','Color','r')

and i get the error given below;

Warning: Failed to converge in 100 iterations during replicate 2.

> In kmeans/loopBody (line 476)

In internal.stats.parallel.smartForReduce (line 136)

In kmeans (line 343)

In tt (line 11)

What may be the problem?

Thank you

[Image Analyst]

You forgot to attach 'ucd1.xlsx', or even any scatterplots. Please do so, so we can help you.

[Mehmet Volkan Ozdogan]

You can find Ucd1 and ucd2.xlsx file in attachment. Thank you