Can Association Rules be Used to Improve Additive Bayesian Network Models?

Abstract

The study of antimicrobial resistance patterns is a challenging task due to the complexity of the mechanisms involved and the potentially large number of relations that must be inferred. Additive Bayesian networks have proved to be a useful tool in this respect, but when data is scarce, modeling results can be improved by integrating complementary information in the learning process. We propose an experimental study to investigate the benefits of including information extracted by association rule mining as prior knowledge into the inference of additive Bayesian networks. We start with an existing additive Bayesian network, generate synthetic datasets from the joint probability distribution that it represents, and attempt to retrieve the original network structure from these datasets by means of model averaging based on a structural Markov Chain Monte Carlo approach, using uninformative priors as well as priors based on the conviction of corresponding association rules. We show that the conviction-guided approach performs consistently better in identifying the true dependency patterns on the simulated datasets.