Bayesian Methodologies Seminar: Bayesian Knowledge Representations & the Fusion of Knowledge from Multiple (Conflicting) Sources

Graphical models that represent knowledge and uncertainty such as Bayesian networks are popular approaches due to their strong probabilistic semantics and oftentimes intuitive decision modeling capabilities. In this talk, we will examine such knowledge representations as Bayesian networks and the more general Bayesian Knowledge-Bases from the point of view of knowledge acquisition and engineering. A primary goal of these models is to provide a powerful, flexible, and semantically meaningful representation that allows the fast and easy direct encoding of knowledge from various information sources - human and otherwise. For example, in the construction of expert systems, the bottleneck lies in the elicitation and encoding of knowledge from the expert. Typically, in a simplistic knowledge representation, the knowledge must be painfully transformed to satisfy the limitations of the representation which often results in information loss. Furthermore, the representation must also be amenable to effective, meaningful, and efficient reasoning. As such, we will also be going further in depth into Bayesian Knowledge-Bases to explore the semantics of its representation, its focus of handling incompleteness, and particularly the issues of knowledge engineering it naturally addresses. With Bayesian Knowledge-Bases, we can further address the challenges of information/knowledge fusion from multiple sources. For example, consider that there are multiple experts building probabilistic models of the same situation and we wish to aggregate the information they provide. There are several problems we may run into by naively merging the information from each - the experts may disagree on the probability of a certain event or they may disagree on the direction of causality between two events (e.g., one thinks A causes B while another thinks B causes A); the experts may even disagree on the entire structure of dependencies among a set of variables in a probabilistic network. While this is problematic for representations such as Bayesian Networks, Bayesian Knowledge-Bases provide a ready and probabilistically sound methodology to accomplish such fusions.