Comprehensibility of Generative vs. Class Discriminative Dynamic Bayesian Multinets

John Burge and Terran Lane

We investigate the comprehensibility of dynamic Bayesian multinets (DBMs) and the dynamic Bayesian networks (DBNs) that compose them. Specifically, we compare the DBM structures resulting from searches employing generative and class discriminative scoring functions. The DBMs are used to model the temporal relationships among RVs and show how the relationships change between different classes of data. We apply our technique to the identification of dynamic relationships among neuroanatomical regions of interest in both healthy and demented elderly patients based on functional magnetic resonance imaging (fMRI) data. The structures resulting from both generative and class discriminative scores were found to be useful by our collaborating neuroscientist, but for differing reasons. For example, generative scores result in structures that illuminate highly likely relationships and are more easily interpreted. Conversely, structures resulting from class discriminating scores are capable of representing more subtle changes and can illuminate important behavioral differences not apparent from structures learned from generative scores.

This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.