[R] predicting fuzzy cluster membership

Prof Brian Ripley ripley at stats.ox.ac.uk
Sat May 24 18:08:14 CEST 2003

That's not how conventional clustering works: you need to understand the 
underlying principles (and for fanny, the main reference).  The goal is to 
place a given set of objects into clusters (overlapping clusters for 
fanny).  It is not to assign a future object to a cluster: that's a 
completely different, supervised, pattern recognition problem, and you 
should be using very different methods (given in different books, even).
Only for a few methods are there closely related prediction methods (e.g. 
1nn on cluster centres for k-means, mda for emclust).

On Sat, 24 May 2003, Luis Torgo wrote:

> I'm trying to obtain a fuzzy clustering with fanny from the cluster package, 
> using a given set of data. That worked just fine.
> I have another separate sample of data from the same problem. For each case in 
> this new sample I would like to know their membership coefficients with 
> respect to the clustering obtained with the first dataset. In effect I want 
> to have a kind of prediction of the probability that each case in the new set 
> belongs to each of the clusters formed with the first set of data. I do not 
> want to add the second ssample to the first and build a new clustering 
> because that would change the initial clustering.
> I've looked in the help pages of the cluster package for some similar example 
> with no success. I've also searched the R mailing list but didn't find any 
> related question.

But it's a question on (lack of) understanding of statistics, and your 
best avenue is to seek expert local statistical help.

Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

More information about the R-help mailing list