An Improved Bayesian Network Super-Structure Evaluation using Physarum polycephalum Bio-inspiration

Abstract

Bayesian Networks are a powerful tool to assess association relationships between variables on a given observed system. Despite their usage on several areas, the task of obtaining such structure exclusively from observational data is a well-documented NP-hard problem. Physarum Learner is a method to obtain a Bayesian Network from a data set, based on the foraging behaviour of the Physarum polycephalum, a member of the Myxomycetes class. However, the original method overlook crucial characteristics of this combinatorial problem. In the work here presented, we implement three changes to the Physarum Learner original algorithm and apply this improved method on three well-known benchmark data sets. Our results indicate a faster convergence point identification, while achieving good structures from a scoring standpoint.