Bayesian Network Construction and Simplified Inference Method Based on Causal Chains

Abstract

A Bayesian network (BN) is a probabilistic graphical model that represents random variables of causal relationships as a directed acyclic graph. There are many methods to construct BNs. These methods decide a BN structure whose likelihood is best in candidates. However, the edges expressing causal relationships tend not to match the one manually obtained by a human, because it reflects the causality between events that do not occur. We should focus on causal relationship of events that occurs in the most of cases. Therefore, it is convenient to generate a BN based on causal chains. To generate a BN from causal chains, we propose an approach to get events and causal chains from diagnostics reports and infer events by using BN. Since causal chains in the report are definitive, probabilities in BNs can be limited to zero or one. Thus, we also propose a simplified algorithm for BN inference.