"BIC selection of the number of clusters in normal mixture models with irrelevant variables", Tomoki Tokuda (University of Leuven, Belgium)

Date: 
2012-08-22

Dear all,

Neural Computation Unit would like to invite you all to a seminar by Dr. Tomoki Tokusa from University of Leuven in Belgium.

Date:  Wednesday August 22nd, 2012
Time:  10:30-11:30
Venue: D-014 (Lab 1 Level D)

Speaker: Dr. Tomoki Tokuda

Title:

BIC selection of the number of clusters in normal mixture models with irrelevant variables

Abstract:

For selecting the number
of clusters in mixture models, the BIC (Bayesian Information Criterion) is
widely used. However, in the presence of irrelevant variables that do not
discriminate between clusters, it is empirically found that the BIC tends to
underestimate the underlying true number of clusters. Yet, the nature of this
problem is not well understood. In this talk, we analytically study the
behavior of BIC in the case of normal mixture models.  First, we derive an analytical approximation
of the expectation of BIC. Using this result, we show that the BIC suffers from
underestimating the true number of clusters in the presence of irrelevant
variables. Second, we propose a solution to this problem in terms of a
corrected BIC. Finally, we report the results of a simulation study, which
gives a first indication that the corrected BIC may have a good performance in
selecting the number of clusters.


Sincerely yours,
Kikuko Matsuo
Administrative Assistant
Doya Unit
Ext. 18689

Location: 
Lab 1, D014
Contact or Sponsor: 
Kikuko Matsuo