Developing A Graph-Based Machine Learning Model For Identifying Money Laundering Networks Associated With Sanctioned Entities In a Bank In Zimbabwe

Abstract

Money laundering involving sanctioned entities presents a critical threat to financial systems, often leveraging complex, hidden transaction networks. This study presents a novel graph-based machine learning approach for detecting money laundering networks (MLNs) linked to sanctioned entities, using banking transaction data. Employing the CRISP-DM methodology, the research systematically progressed through requirement analysis, data engineering, model development, and evaluation to ensure methodological soundness and practical relevance. A Graph Convolutional Network (GCN) was trained on this data, capturing structural dependencies often missed by traditional models. Experimental results demonstrated that the GCN significantly outperformed benchmark classifiers by achieving the highest accuracy of 88.18%, precision of 0.6440, recall of 0.6020, F1 score of 0.7345, ROC AUC of 0.8968, and Matthews Correlation Coefficient (MCC) of 0.7812, offering the best trade-off between sensitivity and specificity. The model effectively balanced detection accuracy and false positive rates, confirming its utility in sanction-linked anti-money laundering (AML) scenarios. This research also validates using key transactional indicators and graph-based learning methods as powerful tools in combating money laundering involving sanctioned entities. Integrating relational data and machine learning advances current AML detection frameworks and supports more effective, data-driven regulatory enforcement. Future studies should use adaptive learning, rectify imbalance, optimize algorithms, incorporate domain data, and evaluate graph-based ML models through practical implementation against changing money laundering strategies.