Graph-Based Feature Learning for Anti-Money Laundering in Cross-Border Transaction Networks

Abstract

The increasing sophistication of cross-border money laundering activities poses significant challenges to traditional detection systems. This paper presents a graph-based feature learning framework specifically designed to identify suspicious transactions in complex, cross-border financial networks. The proposed approach constructs heterogeneous transaction graphs that incorporate multiple entity types and relationship patterns, then applies advanced graph neural architectures to automatically learn discriminative features that capture both spatial transaction structures and temporal behavioral dynamics. The framework addresses critical challenges, including extreme class imbalance, high false positive rates, and multi-jurisdictional complexity through specialized feature extraction mechanisms and adaptive learning strategies. An experimental evaluation of the Elliptic Bitcoin dataset and the IBM IT-AML synthetic banking dataset demonstrates substantial improvements over traditional handcrafted features and baseline graph methods, achieving an 87.3% F1-score with a 73% reduction in false positive rates (from 38.2% to 10.3%). The learned features reveal interpretable patterns in cross-border layering schemes and currency exchange anomalies, providing actionable insights for financial crime investigators.