On identifying significant edges for structure learning in Bayesian networks

Abstract

Bayesian networks (BNs) are widely used as a data analysis tool in diverse areas, including finance, medicine, and sports. A standard data analysis methodology is to use the well-known score-and-search approach to learn a set of possible Bayesian networks and then to perform model averaging with thresholding to identify features such as edges between variables with high confidence. A fundamental step in the methodology is to select the threshold, as the value selected has broad implications for the success of the analysis. However, the problem of selecting a good threshold in Bayesian network structure learning has received limited attention in the literature. In this paper, we identifying an important shortcoming in a widely used threshold selection method. We then propose a simple transfer learning approach for maximizing target metrics and selecting a threshold that can be generalized from proxy datasets to the target dataset and show on an extensive set of benchmarks that it can perform significantly better than previous approaches.