Explainable Anti-Money Laundering

Abstract

Financial institutions face significant challenges in detecting suspicious money laundering activities due to high false positive rates generated by traditional models, leading to substantial manual intervention and high operational costs. Deep learning models, particularly graph neural networks (GNNs), have demonstrated potential to improve detection accuracy. However, the complexity and opacity of GNNs pose interpretability issues, limiting their applicability in the anti-money laundering (AML) context. This thesis explores the application of explainable artificial intelligence (XAI) techniques to enhance the interpretability of GNNs in AML. Using synthetic transaction data based on IBM’s AMLSim project, a GNN model was explained using three XAI methods: LIME, SHAP, and GraphSVX. The evaluation of the XAI techniques focused on five Co12 metrics: coherence, correctness, completeness, compactness, and confidence. When measuring the coherence between the explainers, we found by comparing the largest 5 to 10 feature importances and the corresponding features about 40% to 50% sign agreement in all pairwise comparisons, except between SHAP and GraphSVX. When comparing the rank, however, the agreement was only about 10% to 20%. The results from measuring the correctness of feature importance were inconclusive, necessitating the establishment of a more suitable method for constructing a feature importance ground truth. This gap presents an opportunity for future research since validation of correctness is critical to build trust in the explanations. The evaluation of the correctness of the node importance obtained from GraphSVX, however, indicate that there is a connection between high node importance and the neighbor also being a money launderer, offering valuable insights for investigators. Furthermore, using 31 features, LIME and SHAP showed, on average, high output completeness for the top 15 feature importances and GraphSVX for the top 20 feature and node importances. Lastly, for the confidence measure, the higher R2 value of the GraphSVX surrogate model over LIME suggests that incorporating comprehensive neighborhood information may be crucial.