Enhancing Money Laundering Detection Addressing Imbalanced Data and Leveraging Typological Features Analysis

This research investigated the most effective combination of data balancing methodologies alongside various money laundering typologies to address imbalanced datasets inherent in the classification of money laundering transactions. Utilizing a synthetic imbalanced dataset sourced from IBM ALM, the study examined the effectiveness of three data balancing techniques: Borderline-SMOTE, SMOTE, and NearMiss. Through a comparative examination, encompassing performance evaluation metrics such as accuracy, recall, and confusion matrix, across prominent classification models including Random Forest (RF), Decision Trees (DT), Naive Bayes (NB), and k-Nearest Neighbors (KNN), the investigation aimed to deepen insights into the impact of these methodologies on imbalanced datasets and to delineate their efficacy. The findings underscored that balancing the dataset via SMOTE, coupled with the utilization of the Random Forest classifier, yielded optimal outcomes, boasting an accuracy rate surpassing 91 % and a recall rate exceeding 64%. Furthermore, the analysis of identified typologies revealed association rules indicating that three connections-fan-in, cycle, and mutual- are linked to the fan-out typology.