基于有向异构图的发票明细税收分类方法

Taxation is the material basis for the survival of the country. In order to accelerate tax modernization and facilitate taxpayers to issue value-added tax invoices in a convenient and standardized manner, the State Administration of Taxation requires taxpayers to select the tax classification corresponding to the invoice details before issuing invoices in the tax control system Thus, improving the accuracy of tax classification has a crucial role in the construction of tax risk indicators and analysis of taxpayer behavior characteristics. This paper proposes a classification model based on the Heterogeneous Directed Graph Attention Network (HDGAT). The model significantly improves the accuracy of the tax classification of invoice details by using the directed information among the invoice details and external information.