An Graph Neural Network and Feature Interaction Based Fraud Detection

Zongli Jiang, Jieru Liang, Nan Ma

Abstract: Modern social networks contain various types of objects and rich interactive information, the effect of traditional detection methods is limited. Many researchers proposed graph neural network methods to solve this problem, but most of them have ignored the heterogeneous nature of social networks. So, we proposed a new model named Multi-view Similarity-based Graph Convolutional Network(MSGCN) to address this problem. Our model adopted a multi-view method to deal with rich interaction and heterogeneous graph data. Firstly, we converted the heterogeneous graph into many single-views by using meta- path. Then, we employed a similarity-based graph convolutional network to learn node representation for every single- view. Finally, the multiple views features were aggregated together through the attention mechanism to generate the final representation. Experimental results show that our MSGCN is better than other network representation learning baselines.