Multi-Range Gated Graph Neural Network for Telecommunication Fraud Detection

With the expansion of the mobile communication technology, telecommunication fraud is increasing dramatically which results in the loss of billions of dollars worldwide every year. In recent years, some detection methods utilize data mining and statistical techniques to detect fraud from large amount of subscriber content data, and some methods transform the social relation into a set of topological features, such as degree, k-core etc. However, both content and relation have not been fully explored for identifying fraudsters. In this paper, we propose the Multi-Range Gated Graph Neural Network (MRG-GNN) for learning latent features of social network. Specifically, we first model a social network as a directed graph where vertices with subscriber features represent subscribers and edges with relational features represent activities between them. Then, a novel method based on efficient short walks and node-merging is proposed to structure the convolutions, and graph convolution block captures content information and relation information between users. The multi-range gated readout operation is proposed to aggregate informative features in multiple nodes and automatically learns the representation of user social network. Finally, experiments on a real-world telecommunication network show that our MRG-GNN achieves the state-of-the-art results.