1
$\begingroup$

What do you say about this plot to find the number of cluster for kmean or kproto for mixed data. Where is the elbow to identify? I would say 5? I have 11 feautures.

enter image description here

$\endgroup$
1
  • $\begingroup$ the "elbow method" and silhouette analysis are two different concepts, check e.g. en.wikipedia.org/wiki/… $\endgroup$ – oW_ Nov 20 '17 at 21:20
1
$\begingroup$

The bumps at 8 and 11 are likely just due to random initialization, and if you rerun with a different random seed, then they will be at a different k.

The elbow argument would probably suggest 3, but it is all but clear. I don't think there is a clear cut, but the values only drop as they would on uniform data.

So most likely, a) your distance function is not good enough, 2) the algorithm does not work on this data, and/or c) this evaluation does not work on this data.

$\endgroup$
1
$\begingroup$

You should select 9 as you can see from plot that for the WSS value there is a dip.

It doesn’t matter if you have 2 features or 9 features or n features. Clustering is on Data present in those features(it might depend on the amount of data).

$\endgroup$
5
  • $\begingroup$ I have a little bit problem to validate such findings. Is it really better to use 11 than 8 where also is a dip? When I run multiple times cluster centers look different every time, so it is hard to Analyse the characteristics within each cluster. Any tips ? $\endgroup$ – Tido Nov 21 '17 at 7:48
  • $\begingroup$ Uhm, high is bad. So 11 was worse than both 10 and 12... are you trying to make him pick the worst k? $\endgroup$ – Has QUIT--Anony-Mousse Nov 21 '17 at 7:52
  • $\begingroup$ @Anony-Mousse what other method can you suggest? $\endgroup$ – Tido Nov 21 '17 at 8:34
  • $\begingroup$ Can you do conclusion about cluster tendency from multiple correspondence analysis? Wenn MCA gives also about 10 dimensions to give 80% of variances it is a hint that you need so many clusters? With 3 dimensions only 20% variance can be explained. So 3 clusters are actually not good enough? $\endgroup$ – Tido Nov 21 '17 at 8:37
  • $\begingroup$ "3 dimensions [...] variance explained" is PCA (meaningful only for continuous variables!), not clustering. You mix up things. Also, forget "clustering tendency". That is a synonym for "not uniform distributed". Starting looking at your data. $\endgroup$ – Has QUIT--Anony-Mousse Nov 21 '17 at 15:42

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.