From Networks to Narratives: Bayes Nets and the Problems of Argumentation

Bayesian Belief Networks (BBNs) are gaining traction in practical fields such as law and medicine. Given this growing relevance, it is imperative to make Bayesian methodologies accessible to professionals in these fields, many of whom might lack formal training in probability calculus. Argumentation offers a promising avenue to achieve this. It serves a dual purpose: (i) generating an explanation of the important reasoning steps that occur in Bayesian inference and (ii) exploring the structure of complex problems, which can help to elicit a BBN representation. Since Bayesian probabilistic inference also provides clear normative criteria for argument quality, there is a tight conceptual connection between the argumentative structure of a problem and its representation as a BBN. The primary challenge is representing the argumentative structure that renders BBN inference transparent to non-experts. Here, we examine algorithmic approaches to extract argument structures from BBNs. We critically review three algorithms - each distinguished by its unique methodology in extracting and evaluating arguments. We show why these algorithms still fall short when it comes to elucidating intricate features of BBNs, such as "explaining away" [44] or other complex interactions between variables. We conclude by diagnosing the core issue and offering a forward-looking suggestion for enhancing representation in future endeavors.