How to decide the number of true clusters using kth nearest neighbor clustering?

[partha1]

I used  proc modeclus code with m=1 and k=2 to 10 by1 that is with different values of k but I could not decide which solution was the best. 

[sbxkoenk · Accepted solution]

Hello,

PROC MODECLUS has a (approximate) nonparametric test for the number of clusters.

Read this:

SAS/STAT 14.1 User's Guide

The MODECLUS Procedure
Details tab: Significance Tests
http://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_modeclus_det...

Examples tab: Example 78.3 Cluster Analysis with Significance Tests
http://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_modeclus_exa...

Read also:
Introduction to Clustering Procedures

The Number of Clusters
http://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_introclus_se...

The last few paragraphs of the latter page deal with nonparametric clustering (and more in particular PROC MODECLUS).

It may be worthwhile to determine the number of clusters by using multiple methods and seeing if they agree.

Good luck,

Koen

[partha1]

Thank you. How to fix appropriate radius(r) for the significance test in Kth nearest neighbor?