Harnessing coloured Petri nets to enhance machine learning:A simulation-based method for healthcare and beyond

Many industries use Machine Learning (ML) techniques to enhance systems’ performance. However, integrating ML into these systems poses challenges, often requiring improved explainability and accuracy. Using formal methods is a potential solution to address these challenges. This paper presents a simulation-based method using Coloured Petri Nets (CPN) to enhance the explainability and accuracy of Decision Tree (DT) and Random Forest (RF) models, which industries such as healthcare widely adopt. Our simulation-based method, named RuleXtract/CPN, provides procedures for the automatic extraction of decision rules from an implemented ML model, the generation of these decision rules into a CPN model, the analysis of the CPN model through simulations, and the adjustment of the CPN model to improve explainability and accuracy. Automating the transformation from DT/RF to a CPN model and the analysis procedures can reduce the time and effort needed for modeling tasks. We used web technologies and the Access/CPN framework to implement the procedures defined in our simulation-based method so that users would not need CPN expertise to generate and simulate models, running them in the background. An experiment with three datasets for COVID-19 and five for Influenza screening shows that applying our simulation-based method results in more explainable models. The experiment also shows improvement in accuracy measures for RF models. For instance, the accuracy of the RF model using the Influenza rapid test balanced dataset increased from 84.02% to 86.34%, and the unbalanced dataset from 84.78% to 87.53%. Our results underscore the importance of eliminating duplicated, poorly generalized, and incorrect rules to improve explainability and accuracy. These findings also emphasize the effectiveness of using CPN to improve the models, paving the way for future research.