An Original Investigation of Heterogeneous Graph Representation Learning for Anti-Money Laundering in Nigerian Financial Networks

Money laundering networks in the Nigerian financial system exploit complex, multi-layered transaction patterns that traditional rule-based AML systems are unable to detect effectively, while existing graph learning approaches have not been adapted to the heterogeneous graph structure of Nigerian financial networks. This study develops and evaluates an original heterogeneous graph representation learning approach for anti-money laundering detection in Nigerian financial networks. A heterogeneous financial graph is constructed from transaction data shared by three Nigerian financial institutions under regulatory research arrangements, modelling six entity types (accounts, devices, IP addresses, merchants, branches, agents) and eight relationship types. An original Heterogeneous Financial Graph Neural Network (HFGNN) architecture is proposed, incorporating a meta-path guided attention mechanism that adaptively weights the relevance of different transaction relationship types for AML classification, and a temporal edge encoding component that captures the time-evolving structure of laundering schemes. The HFGNN is compared against homogeneous GNN baselines and conventional rule-based systems on a labelled dataset of 1.2 million transactions. HFGNN achieves an AUC-ROC of 0.964 and an average precision of 0.891, compared to 0.912 and 0.823 for the best homogeneous baseline. Shell company detection, a historically challenging AML task, shows the greatest improvement (F1 improvement of 12 percent). The study constitutes an original contribution to both graph learning theory and Nigerian financial crime detection.

Keywords: graph neural networks, anti-money laundering, Nigeria, financial crime detection, heterogeneous graph learning