Introduction
Traditional tax auditing focuses on individual entities. However, sophisticated evaders utilize "Affiliated-Transaction-Based" (ATTE) strategies, where a web of companies connected by family, ownership, or investment ties shifts profits to dodge taxes. These transactions often appear legal on the surface, making them difficult to detect using standard auditing techniques. Furthermore, even when machine learning identifies suspicious patterns, the lack of explainability hinders official investigations. ATTENet addresses these challenges by combining automated detection with interactive visualization to bridge the gap between high-accuracy detection and human-interpretable auditing.
Methodology
The ATTENet pipeline consists of three main stages: network construction, group detection, and visual analytics.
1. TPIIN Construction
The system first builds a Taxpayer Interest Interacted Network (TPIIN), a heterogeneous graph where nodes represent individuals or companies. Edges represent four critical relationship types: family ties, corporate control, investment, and trade.
2. Suspicious Group Detection
The detection process involves:
- Partitioning: The TPIIN is divided into "tax groups" based on non-trading relationships (e.g., family or ownership).
- Feature Extraction: For each group, basic features (capital, registration time) and structural features are extracted.
- Graph Embedding: The system employs structure2Vec to encode the complex topology of each group into a numerical vector, capturing "suspicious shapes" that traditional statistics might miss.
- Classification: A Random Forest model processes these vectors to generate a suspicion score.
3. Visual Analytics
To provide explainability, ATTENet offers three coordinated views:
- TPIIN Overview: A macro-level view of the taxpayer network.
- Explanation View: Highlights abnormal relationships and trading patterns within a flagged group.
- Statistics View: Details tax burden levels and variations for individual companies.
Data and Results
The system was evaluated using two years of real-world tax data provided by Chinese authorities, encompassing millions of companies and billions of transactions. Using historical tax evasion records as ground truth, the system demonstrated high efficacy in identifying ATTE groups. Qualitatively, ATTENet successfully identified coordinated "tax burden variations" within groups. The visual interface significantly reduced the time required for manual discovery by allowing auditors to immediately see the control and investment edges linking suspicious entities, mapping abstract mathematical values back to visual representations of family trees and trade flows.
Discussion and Future Work
While effective, ATTENet faces challenges in scalability, particularly in partitioning millions of nodes in real-time. Future work will focus on more efficient partitioning algorithms, handling dynamic network changes to spot evasion as it evolves, and incorporating more nuanced relationship types beyond the current four categories.
Summary updated