|
Activity Number:
|
384
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 1, 2007 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #308753 |
|
Title:
|
Structural Learning of Chain Graphs via Decomposition
|
|
Author(s):
|
Zongming Ma*+ and Xianchao Xie and Zhi Geng
|
|
Companies:
|
Stanford University and Harvard University and Peking University
|
|
Address:
|
Sequoia Hall 390 Serra Mall, Stanford, CA, 94305,
|
|
Keywords:
|
Bayesian network ; chain graph ; directed acyclic graph ; junction tree ; structural learning ; undirected graph
|
|
Abstract:
|
Chain graphs form a broad class of graphical models for description of conditional independence structures, including both undirected graphs and directed acyclic graphs as special cases. In this paper, we propose that the structural learning of a chain graph can be decomposed into local structural learning related to its decomposed subgraphs. Algorithms for both skeleton recovery and complex arrow orientation are presented on the basis of the decomposition. The decomposition requires conditional independencies but doesn't require the separators to be complete subgraphs. By decomposition, we localize the search for c-separators into small subgraphs, which improves both the efficiency of structural learning and the power of conditional independence tests.
|
