This research investigates the application of Graph Neural Networks (GNNs) as a novel approach to address the complex challenge of anti-money laundering (AML). By leveraging the inherent graph structure of financial transaction networks, GNNs offer the potential to identify intricate patterns indicative of fraudulent activities that traditional machine learning methods may overlook. The study focuses on a comparative analysis of four prominent GNN architectures—AGIN, GIN, Graph SAGE, and GAT—to assess their efficacy in discriminating between legitimate and fraudulent transactions. A real-world AML dataset was employed to train and evaluate the GNN models. The results demonstrate the superior performance of the AGIN model, achieving a notable accuracy of 0.9215, precision of 0.9348, recall of 0.9768, and AUC of 0.7787, significantly outperforming the other models. To enhance model performance, graph-based features, including node-level and graph-level metrics, were extracted and integrated into the feature space. These features provided valuable insights into the structural characteristics of the transaction network, enabling the models to capture complex relationships and patterns. To address the imbalanced nature of AML datasets, the NearMiss under sampling technique was employed to mitigate the impact of class imbalance on model performance. By balancing the distribution of fraudulent and non-fraudulent transactions, the study aimed to improve the models' ability to accurately identify suspicious activities. The findings of this research contribute significantly to the advancement of AML detection technologies. By demonstrating the effectiveness of GNNs in capturing complex patterns within financial transaction networks, this study offers valuable insights for financial institutions, law enforcement agencies, and policymakers in their efforts to combat money laundering. The proposed framework provides a foundation for future research and development in this critical area.
