Abstract
Anomaly detection in network science is the method to determine aberrant edges, nodes, subgraphs or other network events. Heterogeneous networks typically contain information going beyond the observed network itself. Value Added Tax (VAT, a tax on goods and services) networks, defined from pairwise interactions of VAT registered taxpayers, are analysed at a population-scale requiring scalable algorithms. By adopting a quantitative understanding of the nature of VAT-anomalies, we define a method that identifies them utilising information from micro-scale, meso-scale and global-scale patterns that can be interpreted, and efficiently implemented, as population-scale network analysis. The proposed method is automatable, and implementable in real time, enabling revenue authorities to prevent large losses of tax revenues through performing early identification of fraud within the VAT system.
Introduction: The "Achilles Heel" of VAT
Value Added Tax (VAT) is a cornerstone of revenue for over 160 countries, yet it possesses a structural vulnerability: the invoice-credit mechanism. In this system, businesses subtract the VAT paid on inputs from the VAT collected on sales, remitting only the difference. Fraudsters exploit this through "Missing Trader" (MT) schemes. In these scenarios, a firm (the MT) collects VAT from a buyer and then vanishes without remitting it to the government, while the buyer—and subsequent entities in the chain—continue to claim tax credits.
Traditional detection relies on manual audits or simple firm-level classifiers. However, these methods suffer from selection bias and ignore the communal nature of fraud. MT fraud is rarely an isolated act; it requires a coordinated network of "conduits," "buffers," and "brokers" to obscure the money trail and maximize illicit gains.
Methodology: Risk-Corrected Network Analysis
The researchers propose a hybrid framework that fuses Network Science with Gradient Boosting to identify these fraudulent communities within massive, high-dimensional transactional datasets. The core innovation is the construction of a risk-corrected Laplacian matrix.
- Network Modeling: The economy is represented as a graph where nodes are businesses and edges are VAT-taxed transactions. To handle population-scale data (hundreds of thousands of nodes), the authors utilize a symmetrized adjacency matrix to simplify spectral decomposition.
- Risk-Correction: Standard spectral clustering focuses solely on graph topology. The authors "nudge" the network's Laplacian matrix by incorporating a term derived from firm-specific covariates (e.g., labor costs, employee counts) processed through an XGBoost classifier. This ensures the algorithm groups nodes based on both their trade relationships and their shared "risky" business profiles.
- Scalable Algorithms:
- NIMAD (Network Informed Multiscale Anomaly Detector): A global approach using Singular Value Decomposition (SVD) to classify individual nodes.
- HACI (Hierarchical Anomalous Cluster Identifier): A local approach using recursive bi-partitioning to find entire fraudulent neighborhoods.
Empirical Validation and Results
The study utilized the universe of Bulgarian administrative VAT data from 2016-2017, encompassing 312,762 VAT-registered taxpayers and 49 covariates per firm. The models were validated against a "risk list" provided by the Bulgarian National Revenue Agency (BNRA).
The results demonstrate the power of data fusion:
- Accuracy Improvement: Incorporating network information led to an approximate 19% improvement in AUC compared to models using only firm-level data.
- Detection of New Fraud: For identifying newly registered fraudulent traders, the NIMAD algorithm achieved an AUC of 0.953, significantly outperforming standard XGBoost (0.802).
- Real-World Application: When the model identified the top 200 highest-risk taxpayers, 50% were confirmed as actual new entries on the BNRA’s high-risk list.
Conclusion
By combining network topology with node-level risk profiles, this approach provides a much clearer signal for fraud detection than either method alone. The use of the Lanczos bidiagonalization algorithm ensures scalability, allowing these tools to run on standard hardware in a few hours. While fraudsters may adapt to these methods, the transition to automated, network-informed auditing represents a significant step forward for revenue authorities in preventing large-scale tax losses.
Summary updated