|
Activity Number:
|
85
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 4, 2003 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Bayesian Stat. Sciences
|
| Abstract - #302339 |
|
Title:
|
Bayesian Unsupervised Clustering for Mixed Data with Missing Observations
|
|
Author(s):
|
Halima Bensmail*+ and Hamparsum Bozdogan
|
|
Companies:
|
University of Tennessee and University of Tennessee
|
|
Address:
|
Statistics Department, Knoxville, TN, 37996-0532,
|
|
Keywords:
|
multidimensional scaling ; cluster analysis ; kernel density ; Bayesian analysis ; entropy
|
|
Abstract:
|
We try to develop and broaden the application areas of a new statistical modeling technology which will cover a research study on unsupervised classification with mixed data using multidimensional scaling and Bayesian approach. We attempt to make a bridge and apply Bayesian multidimensional scaling for unsupervised clustering. This new methodology is formulated in the multivariate non-Gaussian settings. We propose a new way of clustering observations described by a mixed collection of variables (nominal, ordinal, and numerical). We then use a nonparametric clustering procedure based on optimally scaling the large data and estimate the distribution of the object scores obtained. We propose a multivariate kernel distribution as nonparametric distribution for the variables, this distribution is characterized by a variety of window widths that control the degree of smoothness of the estimate and a variety of robust covariance matrices. This gives a variety of flexible models to choose from. The prior choice here is challenging especially when dealing with the window width distribution for that we use some conjugate and entropy-based priors to overcome this difficulties.
|
