Integrating machine learning and rule-based systems for fraud detection : a case study based on the logic learning machine

Money laundering is one of the most relevant global challenges, with significant repercussions on the economy and international security. Identifying suspicious transactions is a key element in the fight against the phenomenon, but the task is extremely complex due to the constant evolution of the strategies adopted by criminals and the great amount of data to be analyzed daily. This study proposes a hybrid method that integrates Machine Learning models with heuristic rules, with the aim of identifying fraudulent transactions more effectively. The dataset used, SAML, includes millions of bank transactions and presents a strong imbalance between classes (fraudulent vs regular transactions). The entire process was carried out through a self-code platform designed to optimize data management, processing and analysis. The heuristic rules were evaluated using the covering and error metrics and then integrated into the Logic Learning Machine (LLM) task. The effectiveness of the approach was verified by comparing two main configurations: one based exclusively on the use of LLM and the other combining LLM and heuristic rules. The results obtained highlight that the integration of heuristic rules improves the performance of the model, confirming the synergy between Machine Learning and expert knowledge. This study confirms the effectiveness of the hybrid approach and emphasizes the importance of the union between automated analysis and human insight to address the challenges posed by money laundering.