Enhanced Temporal Graph Embedding for Identifying Fraudulent Transactions on Transaction Networks

Electronic payments have penetrated into every aspect of our lives, and the transaction records of users form transaction networks. Fraudulent transactions in the transaction network bring huge losses to the society, and it is important to identify fraudulent transactions among them. Most existing methods focus on anomaly detection of static graphs, ignoring the information of temporal changes of graphs and transaction payment scenarios. In this paper, we propose a missing prediction dynamic graph anomaly transaction detection model on transaction networks. It consists of three parts: (1) feature extraction layer, in which single evaluation features and common evaluation features are extracted from the ego-net of nodes, and labels of edges are directly used for embedding vector generation. (2) missing information prediction layer, which uses domain models to predict missing information in the temporal graph and complements node features. (3) network evolution layer, which uses GAT in GRU to aggregate graph structure information and to evolve missing node information from both forward and reverse directions. Anomalous edge classification experiments were conducted on three real datasets, and the proposed method outperformed the compared baseline methods.