Identifying synergistic relationships in Bayesian networks: From one-way to two-way sensitivity analysis

Abstract

Bayesian networks are widely valued in artificial intelligence for their capacity to provide interpretable, probabilistic models of complex systems, particularly in settings marked by uncertainty. Robustness and explainability are critical for the adoption of such models in real-world applications, motivating the need to understand how model outputs respond to variations in underlying parameters. While one-way sensitivity analysis is well understood, it does not capture the interactions between parameters that can arise in real-world applications. This thesis addresses the gap in higher-order sensitivity analysis by exploring and classifying the shapes of two-way sensitivity functions in Bayesian networks. We propose a heuristic for the selection of paramter pairs for study and develop and implement an algorithm for calculating two-way sensitivity functions, as well as plot the said fuctions. Using this algorithm we aim to find an efficient way to identify parameter pairs that are likely to have synergistic relationship, using information only from one way analysis.