An edge feature aware heterogeneous graph neural network model to support tax evasion detection

Tax evasion is an illegal activity that causes severe losses of government revenues and disturbs the economic order. To alleviate this problem, decision support systems that enable tax authorities to detect tax evasion efficiently have been proposed. Recent researches tend to use graph to model the tax scenario and leverage graph mining techniques to conduct tax evasion detection, as so to make full use of the rich interactive information between taxpayers and improve the detection performance. However, a more favorable graph mining solution, graph neural networks, has not yet been thoroughly investigated in such settings, leaving space for further improvement. Therefore, in this paper, we propose a novel graph neural network model, named Eagle, to detect tax evasion under the heterogeneous graph. Specifically, based on the guidance of our designed metapaths, Eagle can extract more comprehensive features through a hierarchical attention mechanism that fully aggregates taxpayers' features with their relations. We evaluate Eagle on real-world tax dataset. The extensive experimental results show that our model performs 15.71% better than state-of-the-art tax evasion detection methods in the classification scenario, while improves 5.22% in the anomaly detection scenario.

Automated summary

This paper proposes a novel graph neural network model, Eagle, for detecting tax evasion in a heterogeneous graph. With the guidance of designed metapaths and a hierarchical attention mechanism, Eagle can extract comprehensive features from taxpayers' features and relations, improving the performance of tax evasion detection.