Activity Number:
|
151
|
Type:
|
Topic Contributed
|
Date/Time:
|
Monday, August 3, 2009 : 10:30 AM to 12:20 PM
|
Sponsor:
|
Section on Bayesian Statistical Science
|
Abstract - #303691 |
Title:
|
Sequential Learning in Dynamic Graphical Models
|
Author(s):
|
Hao Wang*+ and Craig Reeson
|
Companies:
|
Duke University and Duke University
|
Address:
|
214 Old Chemistry BLDG, Durham, NC, 27708,
|
Keywords:
|
Bayesian forecasting ; Dynamic matrix-variate graphical models ; Gaussian graphical models ; Loss function ; Multi-process modelling ; Parallel computing
|
Abstract:
|
We propose a natural generalization of the dynamic matrix-variate graphical model (Carvalho and West 2007) to time varying graphs. The generalization uses the multi-process modeling idea to introduce sequential graphical model selection procedures that address uncertainty about graphs while allowing for efficient on-line updates. When comparing models, we propose a predictive criterion where the goal is good prediction by minimizing posterior predictive loss for a given graph, and then, sequentially search for graphs which minimize this criterion. To develop an efficient Bayesian approach for high-dimensional graphical model search, we use a particle stochastic search (PSS) algorithm. The PSS algorithm allows parallel exploration of the search space to find the minima. The model is illustrated using financial time series for predictive portfolio analysis.
|