A network and machine learning approach to detect Value Added Tax fraud

Angelos Alexopoulos, Πέτρος Δελλαπόρτας, Stanley B. Gyoshev, Christos Kotsogiannis, Sofia C. Olhede, Trifon PavkovarXiv 2021 · DOI · arXiv

TL;DR

Researchers have developed a scalable way to catch Value Added Tax (VAT) fraudsters by treating business transactions as a massive social network. By combining graph theory (specifically a "risk-corrected" Laplacian matrix) with machine learning, their approach identifies coordinated fraud schemes that traditional audits miss, improving detection accuracy by nearly 20%.

What Problem Does This Paper Address?

Value Added Tax (VAT) is the "Achilles heel" of many tax systems. It relies on an "invoice-credit" mechanism: businesses pay VAT on their sales but get a credit for the VAT they paid on their inputs. Fraudsters exploit this through "Missing Trader" (MT) schemes. A "missing trader" collects VAT from a buyer and then vanishes without remitting it to the government, while the buyer (and subsequent "buffers" in the chain) still claims the tax credit.

Traditional detection methods struggle because:

  1. Coordination: Fraud isn't an isolated event; it’s a team sport involving networks of firms.
  2. Label Bias: We only know about the fraudsters we’ve caught. Audits are often biased toward certain industries or sizes, making standard supervised learning (like "predicting who is a fraudster based on past audits") unreliable.
  3. Scale: With hundreds of thousands of businesses and millions of transactions, the data is too high-dimensional for simple analysis.

Methods

The authors propose a hybrid approach that merges Network Science with Gradient Boosting.

The core innovation is the Risk-Informed Network Laplacian. In graph theory, the Laplacian matrix is a mathematical representation of a network's structure, often used to find clusters. The authors "corrected" the standard Laplacian by adding a term that accounts for individual firm risk.

They developed two primary algorithms:

  • NIMAD (Network Informed Multiscale Anomaly Detector): This uses global spectral decomposition. It maps the entire network into a low-dimensional space (embeddings) and then uses XGBoost to classify nodes as high or low risk.
  • HACI (Hierarchical Anomalous Cluster Identifier): This is a "cluster-oriented" approach. It recursively splits the network into smaller and smaller sub-groups (communities) to find tight-knit clusters of suspicious activity.

What’s clever here is the scalability. Instead of using a massive matrix of every business characteristic, they first use XGBoost to condense those characteristics into a single "fraud probability" score per firm, which then informs the network model.

Datasets

The study uses the "universe" of Bulgarian administrative data from 2016 and 2017:

  • Population: 312,762 VAT-registered taxpayers.
  • Features: Monthly VAT returns, ledger data (all B2B transactions), number of employees, labor costs, and other risk-based criteria used by the Bulgarian National Revenue Agency (BNRA).
  • Validation: A "toy" simulated dataset was used to prove the math works before applying it to the real-world Bulgarian data.

Key Insights and Ideas

  • Network Structure is a Force Multiplier: Simply adding network features (like how a firm is connected to others) to a standard machine learning model improved the Area Under the Curve (AUC) by 19%.
  • Catching the "New" Fraudsters: The model was exceptionally good at identifying newly registered risky taxpayers—those who hadn't been flagged by previous human-driven systems. In this "detect new" scenario, the AUC jumped from 0.802 (standard ML) to 0.953 (network-informed ML).
  • Symmetry Matters (Sometimes): Even though business transactions are directed (money flows from A to B), the authors found that treating the network as undirected for the Laplacian calculation helped group nodes that share similar "neighborhoods," which is a hallmark of coordinated fraud.
  • Automation vs. Human Bias: By automating the process, the system reduces the administrative cost of selection and avoids the "selection bias" inherent in human-led audit choices.

Key Results

The results are striking for a real-world application:

  • 50% Detection Rate: The method successfully identifies roughly half of the VAT fraud in the Bulgarian system.
  • Efficiency: Using Algorithm 1, the researchers identified 200 taxpayers with the highest fraud probabilities; 100 of these were confirmed as "high-risk" for the first time by the BNRA.
  • Speed: Despite the massive dataset, Algorithm 2 (HACI) completed its analysis in under an hour on a standard laptop, proving it is ready for production environments.

Limitations and Future Work

The authors acknowledge that their current models are static. They look at a snapshot of time (or a series of snapshots), but fraudsters are adaptive—they learn how the tax authorities operate and change their patterns.

Future directions include: * Multi-layer Networks: Incorporating other relationships, such as shared board members or physical addresses, to find deeper connections. * Temporal Dynamics: Moving from static networks to "dynamic" or "inter-temporal" models that track how fraud rings evolve over months or years. * Cross-Border Fraud: Expanding the model to handle intra-community (EU) transactions, where "Carousel Fraud" often crosses national boundaries to further obscure the paper trail.

Summary updated

Abstract

Value Added Tax (VAT) fraud erodes public revenue and puts legitimate businesses at a disadvantaged position thereby impacting inequality. Identifying and combating VAT fraud before it occurs is therefore important for welfare. This paper proposes flexible machine learning algorithms which detect fraudulent transactions, utilising the information provided by the complex VAT network structure of a large dimension. VAT fraud detection is implemented through a combination of a suitably constructed Laplacian matrix with classification algorithms that rely on scalable machine learning techniques. The method is implemented on the universe of Bulgarian VAT data and detects around 50 percent of the VAT fraud, outperforming well-known techniques that ignore the information provided by the network of VAT transactions. Importantly, the proposed methods are automated, and can be implemented following the taxpayers submission of their VAT returns. This allows tax revenue authorities to prevent large losses of tax revenues through performing early identification of fraud between business-to-business transactions within the VAT system.
BibTeX
@misc{alexopoulos2021network,
  author = {Angelos Alexopoulos and Πέτρος Δελλαπόρτας and Stanley B. Gyoshev and Christos Kotsogiannis and Sofia C. Olhede and Trifon Pavkov},
  title = {{A network and machine learning approach to detect Value Added Tax fraud}},
  year = {2021},
  eprint = {2106.14005},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.14005},
}

Loading related content...

Figure