0
$\begingroup$

I have been performing fuzzy c means clustering using Matlab toolbox for my clustering problem. Unfortunately it leads to unstable performance since the selection of parameter membership (Uij) is chosen randomly and the initial centres are derived from Uij. Is there any initial algorithm which can be performed for fuzzy c means algorithm? Thanks

$\endgroup$
2
  • 1
    $\begingroup$ If you get substantially different clusterings w/ different initial centers, that is something to pay close attention to, not something to wave away & ignore w/ some fancy code. $\endgroup$ Apr 26, 2016 at 15:24
  • $\begingroup$ Hi, thanks for your response. So from your statement I can conclude that random initial centers is essential for clustering process since it is a proof that our algorithm is converging or not. Is that right? and also how can we know that the algorithm works or not? I am new at this subject. $\endgroup$
    – M.Faa
    Apr 30, 2016 at 14:50

1 Answer 1

1
$\begingroup$

I agree with the comment by @gung:

If you get very different results every time (not just a permutation of the clusters!) then this indicates the algorithm did not work. It did not converge to a solution that is clearly superior to the others.

You will need to carefully study your result and check if you can preprocess your data better, for example. (In fact, you should even do this if it had converged reliably...)

$\endgroup$
2
  • $\begingroup$ hi, in which condition can we say that our algorithm converges to a solution? Can we see it from the objective function of fcm? $\endgroup$
    – M.Faa
    Apr 30, 2016 at 15:00
  • $\begingroup$ The objective function value is very much data set dependent. You cannot compare it or tell if the value is good. You need to study (!) the actual clutering result. $\endgroup$ Apr 30, 2016 at 15:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.