Semantically-Linked Bayesian Networks: A Framework for Probabilistic Inference Over Multiple Bayesian Networks

At the present time, Bayesian networks (BNs), presumably the most popular uncertainty inference framework, are still widely used as standalone systems. When the problem itself is distributed, domain knowledge has to be centralized and unified before a single BN can be created. Alternatively, separate BNs describing related sub-domains or different aspects of the same domain may be created, but it is difficult to combine them for problem solving even if the interdependent relations between variables across these BNs are available. Existing approaches have greatly restricted expressiveness and applicability as they either impose very strong constraints on the distributed domain knowledge or only focus on a specific application. What is missing is a principled framework that can support probabilistic inference over separately developed BNs.

In this thesis, we propose a theoretical framework, named Semantically-Linked Bayesian Networks (SLBN), to fill this blank. SLBN is distinguished from existing work in that it defines linkages between semantically similar variables and probabilistic influences are carried by variable linkage from one BN to another by soft evidences and virtual evidences. To support SLBN’s inference, we have developed two algorithms for belief update with soft evidences. Both of these algorithms have clear computational and practical advantages over the methods proposed by others in the past. To justify SLBN’s inference process, we propose J-graph to represent the jointed knowledge of the linked BNs and the variable linkages. Finally, SLBN is applied to the problem of concept mapping between semantic web ontologies.