TaxVis: a Visual System for Detecting Tax Evasion Group

Abstract

The demo presents TaxVis, a visual detection system for tax auditor. The system supports tax evasion group detection based on a two-phase detection approach. Different from the pattern matching based methods, this two-phase method can analyze the suspicious groups automatically without artificial extraction of tax evasion patterns. In the first phase, we use a network embedding method node2vec to learn representations that embed corporations from a Corporation Associated Network (CANet), and use LightGBM to calculate a suspicious score for each corporation. In the second phase, the system use three detection rules to analyze the transaction anomaly around the suspicious corporations. According to these transaction anomalies, we can discover potential suspicious tax evasion groups. We demonstrate TaxVis on tax data of Shaanxi province in China to verify the usefulness of the system.