ChatGPT as a Subject Matter Expert in the Parameterization of Bayesian Network Classifiers

• Published in: 2025 28th International Conference on Information Fusion (FUSION) pp. 1 - 8
• Main Author: Kruger, Max J.
• Format: Conference Proceeding
• Language: English
• Published: ISIF 07.07.2025
• Subjects: access to operational knowledge; air surveillance systems; Bayes methods; Bayesian Networks; Chatbots; ChatGPT; classification; classification evaluation; Filling; Jensen-Shannon divergence; Knowledge engineering; Learning systems; parameterization support; parametrization of Bayesian Networks; Subject matter experts; substitution of domain experts; Surveillance
• DOI: 10.23919/FUSION65864.2025.11124000

Bayesian Networks are a well-established approach to solving classification problems in practical applications. Preparing to use a Bayesian Network Classifier requires parameterization of the Bayesian Network, in which the conditional probability tables of the nodes must be filled with probability values that describe the situation and also the dependencies of this node on other nodes. Various parameter learning methods can determine these conditional probabilities from sample data sets. However, in the applications we have in mind, sample data may not exist, for example, due to a change in the application context. In these cases, the parameterization is traditionally carried out with the help of a subject matter expert. We consider situations in which even such a domain expert is unavailable and ask whether chatbots based on LLMs, here ChatGPT, can take over the filling of the conditional probability tables. To answer this question, we consider an operational reference scenario from airspace surveillance for which we have an expert-generated parameterization. On this basis, ChatGPT generates the parameterizations for a Naïve Bayes and a more general Bayes Network Classifier, which are then compared in terms of their classification performance with respect to the reference classifier and against each other.