Counterfactual Methods for Detecting Unfairness in Anti-Money Laundering Algorithms

Anti-Money Laundering (AML) systems frequently incorporate sensitive customer information, such as country of origin, to improve detection performance. This creates a tension between regulatory requirements for risk-based monitoring and the need to ensure algorithmic fairness. The authors investigate whether machine learning models—ranging from tabular decision trees to state-of-the-art graph neural networks (GNNs)—rely on protected attributes in ways that are unfair, specifically by decomposing the causal influence of these attributes into legitimate (mediated) and illegitimate (direct) pathways.

Using the synthetic IBM AMLSim dataset, the authors introduce pseudo-KYC features (sender and receiver country) and a behavioral mediator (transaction volume and frequency). They employ a structural causal model to perform a counterfactual fairness analysis. This allows them to calculate the total effect of the protected attribute on the model's flagging decision and decompose it into a natural direct effect (the potentially unfair component) and a natural indirect effect (the component transmitted through behavioral features, which is considered more acceptable). They compare three models: XGBoost, GNN-GIN, and GNN-PNA.

All models showed improved predictive performance when augmented with country and behavioral features. However, the analysis revealed a clear accuracy-fairness trade-off: the model that benefited most from these features (GNN-GIN) also exhibited the most significant fairness violations, characterized by larger direct effects of the protected attribute. Conversely, the best-performing model (GNN-PNA) showed the smallest causal effects, suggesting it captured relevant patterns without relying heavily on the protected attribute. The study highlights that aggregate fairness metrics can be misleading, as they may obscure the underlying causal mechanisms driving bias.