|
Activity Number:
|
132
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 3, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #304321 |
|
Title:
|
A Comparison of Minimum Description Length Scores for Bayesian Networks
|
|
Author(s):
|
Jason LaCombe*+ and Anthony Almudevar
|
|
Companies:
|
University of Rochester Medical Center and University of Rochester Medical Center
|
|
Address:
|
Dept. of Biostatistics and Computational Biology, Rochester, NY, 14642,
|
|
Keywords:
|
Pedigree Estimation ; Graphical Models ; Bayesian Inference ; Minimum Description Length
|
|
Abstract:
|
The use of Bayesian networks finds applications in many areas of statistical modeling, one example being that of pedigree estimation. A pedigree may be modeled as a Bayesian network with each node possessing at most two parent nodes. The focus of the inference is on correctly estimating the underlying structure of the associated graph. However, an adequate solution has not yet been proposed to address the tendency of maximum likelihood estimation to over-fit the data when many nodes have parents outside the sample. To resolve this problem we use inference methods based on the Minimum Description Length principle. The focus of our work is on comparing the Normalized Maximum Likelihood (NML) score with alternative scoring methods that encode the underlying graph directly. The performance of the methods will be illustrated on simulated pedigrees.
|
