|
Activity Number:
|
399
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 6, 2008 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #300648 |
|
Title:
|
Partially Bayesian Variable Selection in Classification Trees
|
|
Author(s):
|
Douglas A. Noe*+ and Xuming He
|
|
Companies:
|
Miami University and University of Illinois at Urbana-Champaign
|
|
Address:
|
Department of Mathematics and Statistics, Oxford, OH, 45056,
|
|
Keywords:
|
feature selection ; expert opinion ; supervised learning
|
|
Abstract:
|
Currently available statistical models for classification may be split into two broad categories: (1) data-driven and (2) theory-based. Tree-structured classifiers such as CART and QUEST are members of the first category. In those algorithms, all predictor variables compete equally for a particular classification task. However, in many cases a subject-area expert is likely to have some qualitative notion about their relative importance. We introduce a partially Bayesian procedure for dynamically incorporating such qualitative expert opinions in the construction of classification trees. Such an algorithm has two potential advantages. First, unnecessary computational activity can be avoided. Second, we reduce the chance that a spurious variable will appear early in the model. Our models are potentially more interpretable and less unstable than those from purely data-driven algorithms.
|
