Research on financial fraud detection method based on graph neural network

With the popularity of electronic payment and online financial services, financial fraud has become increasingly complex and frequent, and traditional detection methods are difficult to deal with. This paper proposes a financial fraud detection method based on graph neural network (GNN). By constructing a heterogeneous graph model, users and transactions are taken as nodes, and complex relationships are captured by metapaths. In this model, the neighbor information is aggregated by the improved GraphSAGE graph, and the dynamic edge weight attention mechanism is introduced to enhance the ability to identify key transaction associations. Aiming at the imbalance of data categories, combining the focus loss function and dynamic resampling technology, the detection performance of the model for a few types of fraud samples is optimized. The experiment is based on the transaction data of a bank for six months. The results show that this model is significantly superior to the traditional machine learning method and the existing GNN model in terms of AUC-ROC, F1-Score, recall rate and other indicators, and the false positive rate is reduced to 0.012, especially in identifying money laundering transactions disguised as normal behavior. This study provides an efficient and robust new method for financial fraud detection, and will further optimize the detection ability of new fraud in the future by combining the timing diagram network.