Metapath-guided graph neural networks for financial fraud detection

Abstract

Financial fraud detection is an important task to ensure the security of financial system. Graph neural networks has shown good results in the field of financial fraud detection. However there are problems of insufficient data mining and category imbalance in heterogeneous graphs of financial transaction networks. Therefore, this paper proposes Metapath Graph neural networks(Metapath-GNN), a graph neural network model based on metapath subgraph, for detecting financial frauds in complex transaction networks and hidden pattern states. Firstly, the subgraph is generated based on predefined metapath patterns by the metapath subgraph generation module. And the node selection is adjusted using the attention mechanism to improve the adaptability to the category imbalance data; then, an aggregation module is utilized to combine the subgraph and full graph information to generate more representative node embeddings. The effective information is fully exploited to enhance the detection performance of the model. Metapath-GNN is extensively evaluated on public datasets YelpChi, Amazon and Elliptic. In addition, for Elliptic, a real-world financial transaction dataset, the data labeling cost is reduced by a semi-supervised learning approach that makes full use of unlabeled data for training. The optimal performance is also achieved in the comparison experiments with the advanced methods. Such as F1-macro, Area Under the Receiver Operating Characteristic Curve(AUC) and Geometric Mean(GMean), by 11.33%, 1.26%, and 7.00% on YelpChi, 1.75%, 1.31% and 1.22% on Amazon, respectively. In Elliptic key indicator F1 improved by 6.78%. In T-Finance key metrics F1 improved by 1.28% and AUC by 3.54%.