Year: 2025
Type: conference-paper
Abstract: With the rapid development of computing technology, especially deep learning technology, there is an increasing potential to solve complex challenges in various fields including tax risk identification. To address the challenges in VAT invoice risk identification, this paper proposes an innovative deep learning model that integrates a transformer, a temporal convolutional network (TCN) and a heterogeneous graph neural network (HGN). The model addresses three major challenges: feature fusion of multidimensional invoice data, modeling of corporate transactional relationships, and rare tax risk pattern recognition. Specifically, the study constructed a corporate transaction network containing three relationship types (sales, purchases, and queues) and designed a multimodal fusion layer for dynamic feature interaction. The experimental results show that the model can effectively identify traditional and emerging tax risk patterns and greatly improve the timeliness of risk warning. Through deployment and testing in real tax environments, the model successfully identified a large number of previously undetected complex fraud cases and reduced tax losses of about 230 million RMB. The study provides valuable technical support for intelligent tax risk identification and is of great practical significance.
Author: Jinghan Zou
Institution: Beijing Jiaotong University
Language: English
Subfield: Economics and Econometrics
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W4412120067