Title: Parallel Adaptive Multi-Objective Evolutionary Learning of Discretized Bayesian Network Classifiers for Clinical Data

Abstract:

Bayesian Networks (BNs) are gaining traction in the realm of explainable AI due to their capacity to provide transparent probabilistic models that facilitate decision support. Baymex, a recently proposed multi-objective evolutionary algorithm, facilitates the learning of discretized BNs by allowing experts to balance various objectives, including likelihood, model complexity, and prior beliefs. Although Baymex has demonstrated superiority over existing state-of-the-art BN learning methods, it has faced two primary limitations: high computational demands and evaluation restricted to synthetic datasets.

To address scalability, this study introduces a parallelization strategy alongside a mechanism designed to adaptively guide the optimization process toward networks that exhibit reduced overfitting. Additionally, we have reconfigured Baymex to function as a BN classifier, employing multi-objective optimization of cross-entropy loss and the BIC complexity term. This adaptation allows for performance evaluation on real-world clinical classification tasks.

Our results indicate speed improvements of more than 54 times when using a 16-core CPU. Furthermore, comparative analyses against clinically established baselines—specifically decision trees, logistic regression, naive Bayes, and random forests—across two public datasets (RADCURE and SUPPORT) and one proprietary dataset, reveal that Baymex achieves predictive performance that is either statistically comparable to or superior to these benchmarks. Crucially, Baymex generates compact, clinically inspectable BNs and identifies multiple plausible classifiers containing predictors that align with recognized clinical factors.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC