Análise formal de sistemas baseados em aprendizado de máquina usando redes de petri coloridas.

Abstract

Machine Learning (ML) has been widely applied in critical areas such as healthcare, man ufacturing, and transportation. However, its integration into critical systems requires greater explainability and accuracy. Models like Decision Trees (DT) and Random Forests (RF) can generate redundant rules, making interpretation difficult and compromising transparency. DTs increase in depth and number of nodes as they capture patterns, which can hinder inter pretation, affecting transparency and applicability in critical systems, particularly in health care. This thesis presents a method based on Coloured Petri Nets (CPN) aimed at improving the explainability of DT and RF models. The method, named RuleXtract/CPN, automates the extraction, analysis, and adjustment of decision rules, allowing these steps to be per formed by users without expertise in CPN. The developed method consists of transforming DT and RF models into CPN models. Through simulations, the decision rules are ana lyzed and adjusted, eliminating redundancies and identifying specific or incorrect rules that produce misleading classifications. The method was implemented using web technologies integrated with the Access/CPN framework, so users do not need CPN expertise to generate and simulate models, executing them in the background. Experiments were conducted with six COVID-19-related datasets and five related to Influenza. The results show a significant reduction in the number of decision rules: in the balanced dataset, the rules were reduced from 882 to 688, while in the imbalanced dataset, the reduction was from 876 to 687. The elimination of redundant rules reduced the complexity of the models, making it easier for experts to validate them before adopting the rules in a Decision Support System (DSS). The f indings highlight the relevance of the method in increasing trust and explainability of ML models applied to critical systems. The developed methodology presents potential for future research, including its scalability and application to other ML algorithms.