Abstract

As artificial intelligence increasingly shapes decisions that affect people’s lives, users need to understand both what a model predicts and why. Bayesian networks (BNs) are transparent probabilistic graphical models, well suited to reasoning under uncertainty, yet their outputs still need post-hoc explanation to be meaningful to end-users. This talk examines explanation of evidence in BNs: determining which unobserved variables, and in which states, best account for the observed evidence. The most probable explanation often yields overly specified answers that are difficult to interpret. The most relevant explanation (MRE) addresses this by favouring concise, relevant subsets of variables, but its computational cost limits its use in practice. We showcase our previously proposed forward-gLasso algorithm, which approximates MRE with accuracy comparable to an exhaustive local forward search while searching a smaller neighbourhood on average. Making explicit the factors that drive a model’s reasoning allows users to identify when sensitive attributes, or proxies for them, are influencing outcomes. In this way, explanation of evidence provides a foundation for fairness and trust in probabilistic decision support.