different output in kmeans

i used kmeans for clustering similar images.... if i run the code first i get the correct clusters..... but without closing matlab if i execute the second time for the same image, it is clustering different output.... why like that..... what shud i do to get the same output whenever i execute the code... please do reply.....

Answer 1

[Youssef Khmou]

hi, i think this question has been asked before, the reason is that the K-means algorithm starts with random partition so every time you run the code, you get the same result but with different RMSE.

(try to clear the Workspace and re-run ...)

[Elysi Cochin]

yes sir i too guessed it may be of the random partitioning of kmeans.... so i cleared all the variables using "clear all" and executed it..... but the same thing.... unless i close matlab and re-run it i'm not getting the same output... sir is there anything i can do to get same output..... please do reply sir....

[Youssef Khmou]

ok try to add the option 'start' as in this page : http://www.mathworks.com/support/solutions/en/data/1-EF8997/index.html?product=ML&solution=1-EF8997

[José-Luis]

@Youssef: no, you don't get the same result, if by result you mean centers. It might happen that you get the same RMSE for different centers. That's rather unlikely though. Or the same centers and different RMSE, but that's also very unlikely.

[Walter Roberson]

http://www.mathworks.co.uk/matlabcentral/answers/74592-k-means-clustering-result-always-changes#answer_84423

[Walter Roberson]

Could you indicate

size(repmat(minc, nsamp, 1))
size( bsxfun(@times, (0:nsamp-1).', (maxc - minc) ./
  (nsamp-1)) )

Answer 2

[José-Luis]

An option is to reset the random number generator to its initial state every time before running your code:

rng default % ->This is the important bit
X = [randn(100,2)+ones(100,2);...
     randn(100,2)-ones(100,2)];
opts = statset('Display','final');

[idx,ctrs] = kmeans(X,2,...
                    'Distance','city',...
                    'Replicates',5,...
                    'Options',opts);

This will always produce the same result, but it sorts of beat the purpose of the function and might produce bad results.

[José-Luis]

For example:

X = [randn(100,2)+ones(100,2);...
     randn(100,2)-ones(100,2)];
opts = statset('Display','final');

[idx,ctrs] = kmeans(X,2,...
                    'Distance','city',...
                    'Replicates',1,...
                    'Options',opts,...
                    'start',[0.25 0.25; 0.75 0.75]);

But that does not guarantee that the result will always be the same.

[Elysi Cochin]

thank you all for your valuable suggestions in helping me to solve my problem..... thank you all once again......