I read on my machine learning course (on coursera) that random initialization performed several times and then taking the cluster with the lowest cose could help when the number of clusters is "small", but didn't help much for K>>10. Why so?

  • $\begingroup$ Random initialization compared to what initialization? $\endgroup$ – ttnphns Feb 25 '14 at 15:13
  • $\begingroup$ @ttnphns I just edited my post. What I mean is that, in the K-means algo I learned, one method consisted of applying k-means several times with different initial centroids and then taking the clustering which gave the lowest cost. However, the course said that this method didn't help much when K was large. I want to know why? $\endgroup$ – bigTree Feb 25 '14 at 15:16
  • 1
    $\begingroup$ I think, given the data of a fixed No of points in a fixed No of dimensions, the greater is K of random initial centres the smaller is the room for a considerably suboptimal solution to occur. After all, if K = N of objects, the "solution" is always the best :-) $\endgroup$ – ttnphns Feb 25 '14 at 15:25
  • $\begingroup$ @ttnphns I like this explanation intuitively. $\endgroup$ – bigTree Feb 25 '14 at 15:32

Your Answer

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

Browse other questions tagged or ask your own question.