Graph Convolutional Networks Detect Suspicious Transaction Patterns in Banking Systems

Abstract

Financial fraud represents a persistent and escalating threat to banking systems worldwide, demanding detection methodologies that can transcend the structural limitations of conventional rule-based engines. This review examines the application of graph convolutional networks (GCNs) to the identification of suspicious transaction patterns in modern banking ecosystems, analyzing how graph-structured representations of transactional data enable models to uncover organized fraud schemes that remain invisible to tabular machine learning (ML) approaches. Key architectural innovations are surveyed, including spectral convolution, graph attention mechanisms, multi-relational graph modeling, and hybrid temporal-structural frameworks, alongside training strategies designed to address the severe class imbalance characteristic of real-world fraud datasets. The review further examines federated learning integration for privacy-preserving collaborative detection, explainability frameworks for regulatory compliance, and benchmark evaluations across publicly available and proprietary financial datasets. Findings consistently demonstrate that GCN-based systems outperform classical ML baselines and offer compelling pathways toward the next generation of anti-money laundering (AML) surveillance infrastructure.