AN EXPLAINABLE GRAPH NEURAL NETWORK FRAMEWORK FOR ANTI–MONEY LAUNDERING IN CRYPTOCURRENCY TRANSACTIONS USING THE ELLIPTIC DATASET

Abstract

The detection of illicit cryptocurrency transactions remains a significant challenge due to the extreme class imbalance and limited generalization capabilities of machine learning models applied to Anti–Money Laundering (AML) data. In the widely used Elliptic dataset, illicit transactions represent less than 2% of all nodes, creating a high-risk setting in which models can achieve deceptively high training accuracy while failing to meaningfully identify malicious behavior. This study examines the behavior of Graph Neural Networks (GNNs) under these constraints and emphasizes the limitations rather than the performance of the approach. Instead of treating the model’s high training accuracy as a success, we demonstrate how imbalance, structural sparsity, and label noise impede reliable learning. We evaluate the model with and without common imbalance-handling strategies including class weighting and focal lossand illustrate that performance remains unstable. Furthermore, we investigate the explainability of the model using GNNExplainer, showing example subgraphs and salient features for known illicit nodes, and discuss their alignment with money-laundering patterns such as fan-out and transaction mixing. Our findings underscore the difficulties of applying GNNs to heavily imbalanced AML datasets and highlight the need for improved modeling strategies, semi-supervised techniques, and more robust explainability methods for real-world financial crime detection.