Activity Number:
|
147
|
Type:
|
Contributed
|
Date/Time:
|
Monday, August 12, 2002 : 2:00 PM to 3:50 PM
|
Sponsor:
|
Section on Bayesian Stat. Sciences*
|
Abstract - #301121 |
Title:
|
Modal Clustering in One-Dimensional, Conjugate Dirichlet Process Mixture Models
|
Author(s):
|
David Dahl*+
|
Affiliation(s):
|
University of Wisconsin, Madison
|
Address:
|
906 Eagle Heights Apt. F, Madison, Wisconsin, 53705, U.S.A.
|
Keywords:
|
Perfect sampling ; Clustering ; Dirichlet process ; Product partition model ; Coupling ; Model-based clustering
|
Abstract:
|
The Dirichlet process mixture (DPM) model is a popular nonparametric Bayesian tool for modeling unknown distributions through mixtures of components. Integrating out the latent location variables leads to a product partition model for clustering observations. This paper describes a dynamic programming algorithm for univariate data which quickly finds either the maximizer of the posterior clustering distribution or the maximizer of the clustering likelihood. Applications of the algorithm are shown. The algorithm plays a key role in a coupling-from-the-past procedure for obtaining exact draws from a conjugate DPM model with a small number of observations. On a much larger scale, the algorithm is used as a clustering technique for a microarray dataset of over 10,000 genes.
|