MATLAB Answers

0

Why kmeans gives different results each time?

Asked by huda nawaf on 18 Dec 2014
Latest activity Commented on by huda nawaf on 19 Dec 2014
* *I have square binary similarity matrix show the social relation among users, where o means no relation between two users and 1 means there is relation between them.
I used kmeans to do clustering*
f1=dlmread('d:\matlab\r2011a\bin\paper_comm\link_flixster_bin1.txt');
c=kmeans(f1,3);
When run the kmeans more than one times, the results are different.
for example at firs time the cluster 1= 4448 users , cluster 2= 434, and cluster 3=118
But, in second times cluster 1= 4880 users , cluster 2= 119, and cluster 3=1
Why the results are different??*

  0 Comments

Sign in to comment.

Tags

2 Answers

Answer by John D'Errico
on 18 Dec 2014
 Accepted Answer

kmeans uses random starting values. (READ THE HELP. I just did to verify this.) So why would you expect that the solution will be identical if the start points are not?

  1 Comment

Thanks,
I forget this information

Sign in to comment.


Answer by Chetan Rawal on 18 Dec 2014

As John mentioned, the clustering happens by starting at random points, automatically selected by the algorithm. That is why in such a optimization/machine learning problems, you should try multiple iterations and use a validation data set if possible. To get the results closer between different runs, you can try to:
  • Increase number of iterations by increasing 'MaxIter'
  • Use your own starting points with the 'start' name-value pair
Starting with your own seeds instead of randomly selected seeds by MATLAB will ensure a consistent answer.

  2 Comments

how start with my seeds? and how set the seed?
thanks

Sign in to comment.