GM-DualTransAML++: A Multi-Source Graph–Transformer Framework with Statistical Risk Mining and Typology Supervision for Anti-Money Laundering Detection

Abstract

Anti-Money Laundering (AML) has become an increasingly critical challenge for financial institutions because of the rapid growth of digital transactions, sophisticated laundering strategies, and the extreme imbalance between legitimate and suspicious activities. Conventional rule-based systems and existing machine-learning models suffer from high false-positive rates, limited interpretability, inadequate modeling of transaction relationships, and poor generalization to complex laundering behaviors. This paper proposes a graph- and data-mining-enhanced dual-encoder Transformer model for large-scale AML transaction detection. The proposed model combines graph-mining features, AML statistical risk estimation, supervised association-rule mining, sequential Transformer-based behavioral modeling, and typology-supervised multi-task learning within a unified end-to-end framework. In addition, focal loss and class-balanced sampling are used to address the severe class imbalance in AML datasets, and the model is evaluated using chronological train-validation-test splitting together with ten-fold forward-chaining cross-validation. The performance of the proposed model is evaluated using the publicly available Synthetic Anti-Money Laundering Dataset (SAML-D). Experimental results show that the proposed model outperforms conventional baseline models. GM-DualTransAML++ achieves an accuracy of 98.86%, precision of 65.70%, recall of 31.73%, F1-score of 42.79%, MCC of 0.4518, Cohen's Kappa of 0.4229, ROC-AUC of 0.9286, and PR-AUC of 0.3127, with a log loss of 0.0947 and Brier score of 0.0119. Ablation and statistical significance analyses further show that graph mining, statistical risk modeling, association-rule extraction, sequential behavioral learning, and typology supervision each contribute to detection performance. The alert-budget evaluation also demonstrates the practical utility of the proposed model for prioritizing high-risk transactions for investigator review.