FCM for supervised learning

Vardhan Walavalkar (vardhan@bu.edu)
Tue, 28 Apr 1998 17:18:03 +0200 (MET DST)

I just kind of finished a class project for which a number of
supervised problems had to be solved using a given network. I chose the
Fuzzy C Means architecture which is basically an unsupervised recognition
task, and used it in a way to 'learn' from a training set, using a 'home
grown' method of forming ever fine clusters until the req accuracy criterion
is met. Unfortunately I couldnt experiment with the diff varieties of FCM
and am still wondering if there is anyone out there who had a chance to try
out such a thing. also is there someone reference for such kind of work?
the only reference I found was on the web:
Knowledge based clustering: http://seraphim.csee.usf.edu/kb-clus.html
where they use a similar method.

Some questions I had were:
a) how are the cluster centers and memberships obtained from the
fuzzy-c-means algorithm for the training samples generalised to any
new pattern (test data)? I used a weighted sum of the memberships for each
training pattern based on inverse distance weighing from the test points.
b) has anyone tried using Gustafson-Kessels shape describing covariance
matrices for clustering? how well does it detect the cluster shapes for
different problems?
c) finally, how well can a FCM algorithm be used for problems
for which there are no inherent well separated clusters in the data, but
some additionnal information (such as a set of supervised training data
with a class assigned to each pattern) is available.

Many thanks for your useful help!



#-----------------#-----------------#------------------#------------------# Vardhan Walavalkar <----- /| 215, Harvard Ave #3, <----- TWaNg! / | Allston, MA02134. <----- \ | http://crs-www.bu.edu/crs/Vardhan/ <----- \| #-----------------#-----------------#------------------#------------------#

This theory is worthless. It is not even wrong! - Wolfgang Pauli