Application of GraphSAGE in Financial Fraud Detection within Dynamic Social Networks

This study explores the application of the GraphSAGE model for financial fraud detection in dynamic social networks. The dataset, sourced from various business periods of enterprises, constructs a fully connected directed dynamic graph of a social network. Nodes represent registered users, while edges denote emergency contact relationships between users and include time and type attributes. Our objective is to improve the accuracy of financial fraud detection by identifying fraudulent users within these graph structures. In the feature engineering phase, we extracted complex features such as node in-degrees and out-degrees, edge types, and time information. We then employed the GraphSAGE model for modeling and prediction. GraphSAGE effectively captures local relationships in social networks by learning and aggregating information from a node’s neighbors, making it suitable for handling dynamic and complex network data. Experimental results show that the standard GraphSAGE model achieved an accuracy of 77.27% in this study, while the GraphSAGE model with neighbor sampling further improved the accuracy to 78.10%. These results validate the effectiveness and superiority of GraphSAGE in handling dynamic social network data, particularly in capturing complex relationships and temporal dynamics between nodes.