Prev: Vector * Matrix
Next: Minimum Cluster Distance for RBF?
From: Fendwick on 29 Mar 2010 11:26
Hi all:

I keep seeing references made to Hierarchical Clustering (HC) when the number of clusters is not known a priori, and much appreciate the helpful online example presented at: http://www.mathworks.com/access/helpdesk/help/toolbox/stats/bq_679x-3.html

I understand that the height of the links is relative to the distinguishability of the associated subclusters, making solutions that cut through long vertical lines (presumably the 'natural divisions') in the dendrogram good solutions.
However despite references made to natural divisions, the only parameters I see being used in the example for the cluster command seem arbitrary: either an inconsistency coefficient cutoff 'c', or the very maxclust 'n' that I'm looking to HC to find in the first place.

While I can see in this simple example that a straight horizontal line might well partition the data along natural divisions, nothing indicates that all the natural divisions in a larger, more complicated example would be so conveniently horizontally aligned. Though some determination could be made visually, I see no option for HC itself to suggest an optimal clustering scheme programmatically -- e.g. by considering the top 'i' inconsistency coefficients.

Is there an option to have HC suggest an optimal clustering scheme without needing to specify a priori either a maxclust 'n' or an inconsistency coefficient cutoff 'c'?
 | 
Pages: 1
Prev: Vector * Matrix
Next: Minimum Cluster Distance for RBF?