A Novel Method for Aggregation of Bayesian Networks without Considering an Ancestral Ordering

Abstract

A good method of combining Bayesian networks (BNs) should be a generic one that ensures a combined BN meets three important criteria of avoiding cycles, preserving conditional independencies, and preserving the characteristics of individual BN parameters. All combination methods assumed that there is an ancestral ordering shared by individual BNs. If this assumption is violated, then avoiding cycles may be inefficient. In this paper, without considering an ancestral ordering, we introduce a novel method for aggregation of BNs. For this purpose, we first combine the BNs using the modification of the method introduced by Feng et al. We then use the simulated annealing algorithm for getting an acyclic graph in which the minimum arcs have been removed. Using this method, most of the conditional independencies are preserved. We compare the results of the proposed method with the two classical BNs combination methods; union and intersection, and hence to demonstrate the distinctive advantages of the proposed BNs combination method.