Network-to-Prior Bayesian Network Learning for Interpretable Risk Decision Support

Abstract

Complex-network analysis describes risk relations, whereas Bayesian networks (BNs) support probabilistic updating and scenario analysis. We propose a network-to-prior interface that converts semantic roles, normalized mutual information, communities, and bridge participation into a feasible edge domain and data-revisable candidate-edge support. The underlying BIC score remains unchanged, so supported edges may still be rejected during search. In independent-sample simulations, directed F1 increased from 0.243 to 0.434 at N = 50, and project-data resampling showed higher edge stability. Component controls indicate that role restrictions and candidate-edge membership produce most of the structural effect, while the 1.0/0.75 tier distinction adds little. Under cross-fitting, predictive gains were uncertain; in service-disjoint Federal IT data, support changed learned structures while predictions remained stable. The method is therefore best viewed as a structural regularizer for inspectable BN search and posterior scenario analysis, rather than a universally superior predictor.