K-means Clustering Result Always Changes

I'm working on k-means in MATLAB. Here are my codes:

load cobat.txt
k=input('Enter the number of cluster: ');
if k<8
    [cidx ctrs]=kmeans(cobat, k, 'dist', 'sqEuclidean');
    Z = [cobat cidx]
else
    h=msgbox('Must be less than eight');
end

"cobat" is the file of mine and here it looks:

65  80  55
45  75  78
36  67  66
65  78  88
79  80  72
77  85  65
76  77  79
65  67  88
85  76  88
56  76  65

My problem is everytime I run the code, it always shows different result, different cluster. How can I keep the clustering result always the same?

Answer 1

[Walter Roberson]

%generate some initial cluster centers according to some deterministic algorithm
%in this case, I construct a space-diagonal equally spaced, but choose your
%own algorithm

minc = min(cobat, 1);
maxc = max(cobat, 1);
nsamp = size(cobat,1);
initialcenters = repmat(minc, nsamp, 1) + bsxfun(@times, (0:nsamp-1).', (maxc - minc) ./ (nsamp-1));

%Once you have constructed the initial centers, cluster using those centers

[cidx ctrs] = kmeans(cobat, k, 'dist', 'sqEuclidean', 'start', initialcenters);

[Alvi Syahrin]

Thank you for the answer.

I still get confused of this line:

initialcenters = repmat(minc, nsamp, 1) + bsxfun(@times, (0:nsamp-1).', (maxc - minc) ./ (nsamp-1));

Do you mean; I have to construct how to initial the central of the cluster in matlab? But it (k-means) picks randomly, no method. Then do you have any idea?

[Walter Roberson]

kmeans does not pick randomly if you pass the argument 'start' and a data matrix of initial cluster centroids.

That particular line constructs evenly-spaced points between the minimum and maximum values of each column -- similar to using linspace() but working on all the columns at once. The details are not really important: I just chose those centroids to have a bunch of centroids that were deterministically distributed around the entire space of values.

[Alvi Syahrin]

Ok. It's been fixed now, by changing some scripts. Thank you for the idea.

[Mohammad Al Nagdawi]

Excuse me! what are the parameters that kmean used to cluster image pixels? an intensity and position of the pixel or something else.

Thanks

[esmat abdallah]

initialcenters = repmat(minc, nsamp, 1) + bsxfun(@times, (0:nsamp-1).', (maxc - minc) ./ (nsamp-1));

please, matlab out an error on this line : "Error using +

Matrix dimensions must agree."

what can i do ??

[Walter Roberson]

@esmat abdallah

%generate some initial cluster centers according to some deterministic algorithm
%in this case, I construct a space-diagonal equally spaced, but choose your
%own algorithm

minc = min(cobat, [], 1);
maxc = max(cobat, [], 1);
nsamp = size(cobat,1);
initialcenters = repmat(minc, nsamp, 1) + bsxfun(@times, (0:nsamp-1).', (maxc - minc) ./ (nsamp-1));

%Once you have constructed the initial centers, cluster using those centers

[cidx ctrs] = kmeans(cobat, k, 'dist', 'sqEuclidean', 'start', initialcenters);

Answer 2

[the cyclist]

K-means clustering uses randomness as part of the algorithm Try setting the seed of the random number generator before you start. If you have a relatively new version of MATLAB, you can do this with the rng() command. Put

rng(1)

at the beginning of your code.

[Alvi Syahrin]

Thank you for the answer. I have MATLAB 7.11.0(R2010b), and when I tried that command, it's not working, getting an error for undefined function. Do you have any idea to solve this?

[the cyclist]

Type

>> doc randstream

to see how to do it in your version.

Answer 3

[Pallavi Saha]

I am facing the same issue inconsistency in the output of fcm. Can anyone help me

[Walter Roberson]

K-means select cluster positions randomly unless you pass it specific start positions.

To stop the output from being random, you can either pass specific start positions for the clusters, or you can use rng() to set the random seed to a constant so that the same sequence of random values will be created each time.

[Mehmet Volkan Ozdogan]

Hi,

I have a question about rng(). If we use rng() command, K-means algortihm stil repeats until the results are getting convergenced to the best. Is that right?

Thank you

[Walter Roberson]

Yes.

rng(SomeParticularNumericSeed)

just ensures that it will always use the same random number sequence provided that no other random numbers are asked for between the rng() call and the kmeans call.