Multi-view graph neural network for fraud detection algorithm

Abstract

Aiming at the problem that in the field of fraud detection, imbalance labels and lack of necessary connections between fraud nodes, resulting in fraud detection tasks not conforming to the hypothesis of homogeneity of graph neural networks, multi-view graph neural network for fraud detection (MGFD) algorithm was proposed.First, A structure-independent encoder was used to encode the attributes of nodes in the network to learn the difference between the fraud node and the normal node.The hierarchical attention mechanism was designed to integrate the multi-view information in the network, and made full use of the interaction information between different perspectives in the network to model the nodes on the basis of learning differences.Then, based on the data imbalance ratio sampled subgraph, the sample was constructed according to the connection characteristics of fraud nodes for classification, which solved the problem of imbalance sample labels.Finally, the prediction label was used to identify whether a node is fraudulent.Experiments on real-world datasets have shown that the MGFD algorithm outperforms the comparison method in the field of graph-based fraud detection.

Keywords

• fraud detection;anomaly detection;attention mechanism;graph representation learning;imbalance learning