I would like to know whether there is any condition that using one cluster evaluation methods (Dunn, Silhouette, Davies-Bouldin) is better than others. I didn't find anything online or inside books, except a short descriptions that Dunn index is good for non-convex shaped clusters, but didn't find a proper justification for this claim.
$\begingroup$
$\endgroup$
-
1$\begingroup$ They will often prefer the same solutions and the differences are supposedly largely a matter of subjective preference. $\endgroup$ – Has QUIT--Anony-Mousse Aug 19 '18 at 9:15
-
$\begingroup$ For non-convex clusters, DBCV may be more appropriate. $\endgroup$ – Has QUIT--Anony-Mousse Aug 19 '18 at 9:19
-
$\begingroup$ Please go to my web-page and read "Clustering criterions" documentation. The three criterions are very different in their "ideology" $\endgroup$ – ttnphns Aug 19 '18 at 10:16
$\begingroup$
$\endgroup$
Bolshakova and Azuaje (2003) presents an evaluation between the three cluster validation techniques you mention. Maybe this helps if you haven't seen it.
-
$\begingroup$ Thank you for the paper. I go thought it, and it is domain specific (genome data) I would like to know something in general. $\endgroup$ – Reza Rawassizadeh Aug 19 '18 at 12:16
-
$\begingroup$ Oh, I see. Sorry that I couldn't help, then. $\endgroup$ – Ozan C. Aug 19 '18 at 21:20
