|
Activity Number:
|
149
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, July 30, 2007 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Nonparametric Statistics
|
| Abstract - #308580 |
|
Title:
|
Clusterwise Regression Using Dirichlet Mixtures
|
|
Author(s):
|
Subhashis Ghoshal*+ and Changku Kang
|
|
Companies:
|
North Carolina State University and Bank of Korea
|
|
Address:
|
2501 Founders Drive, Raleigh, NC, 27695-8203,
|
|
Keywords:
|
Nonparametric regression ; clustering ; Dirichlet mixture process ; MCMC ; curse of dimensionality
|
|
Abstract:
|
We consider a clustering based approach to non-parametric regression when data come from finitely many hidden sub-populations in each of which a simple parametric regression model holds. We try to recover the lost labels by a Bayesian clustering technique based on Dirichlet mixture of normals. Because sample is not split regionwise, we avoid the curse of dimensionality problem in higher dimension. The clusters are formed automatically within an MCMC scheme. Model parameters are estimated by least square method in each cluster. An ensemble of parametric regression estimates are formed, each based on a configuration formed in each MCMC step, and a simple averaging produces the final estimate. Our method also gives confidence bands and compares favorably with kernel, spline or GAM based method when clusters overlap less. The method is applied to analyze a donation data.
|
