[Paper Review] A network and machine learning approach to detect Value Added Tax fraud

Angelos Alexopoulos, Πέτρος Δελλαπόρτας+4 more|arXiv (Cornell University)|Jun 26, 2021

Imbalanced Data Classification Techniques4 citations

TL;DR

This paper proposes a novel machine learning framework that integrates network science with scalable classification algorithms to detect Value Added Tax (VAT) fraud by analyzing the structural patterns of B2B transaction networks. Using Bulgarian VAT data, the method detects approximately 50% of known fraud cases—significantly outperforming traditional techniques that ignore network structure—enabling early, automated fraud identification at scale.

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.

Motivation & Objective

• To develop an automated, scalable fraud detection system for Value Added Tax (VAT) that leverages the structural properties of business transaction networks.
• To improve fraud detection accuracy by integrating taxpayer-level behavior with group-level network patterns, overcoming limitations of traditional methods.
• To enable early identification of fraudulent taxpayers during the VAT return submission process, reducing revenue loss before it occurs.
• To provide a practical, deployable solution for tax authorities to prioritize high-risk taxpayers for auditing using minimal manual intervention.
• To address the inter-temporal and dynamic nature of fraud by proposing a foundation for future multi-layer network models.

Proposed method

• Constructs a Laplacian matrix from the VAT transaction network to encode structural relationships among businesses.
• Applies scalable machine learning classifiers (e.g., random forests, gradient boosting) on features derived from the network Laplacian and individual taxpayer behavior.
• Uses a two-stage algorithm: Algorithm 1 identifies high-risk taxpayers based on network and behavioral features; Algorithm 2 refines the ranking using additional statistical filtering.
• Employs a risk-list generation mechanism that ranks taxpayers by predicted fraud probability, allowing authorities to prioritize audits.
• Validates performance using real-world data from the Bulgarian National Revenue Agency, comparing detection rates against baseline methods.
• Incorporates network topology features such as centrality, clustering, and community structure to detect abnormal transaction patterns.

Experimental results

Research questions

• RQ1 How can network structure in B2B VAT transactions be leveraged to improve fraud detection beyond individual taxpayer behavior?
• RQ2 Can machine learning models trained on network-embedded features detect a higher proportion of actual VAT fraud than conventional methods?
• RQ3 What is the optimal trade-off between false positive rate and detection rate when selecting taxpayers for audit using automated classification?
• RQ4 How does the integration of network science with machine learning enhance the scalability and automation of fraud detection in large-scale tax systems?
• RQ5 To what extent can the proposed method identify fraud before it causes significant revenue loss?

Key findings

• The proposed method detects approximately 50% of known VAT fraud cases in the Bulgarian dataset, significantly outperforming traditional techniques that do not use network structure.
• Algorithm 1 achieves superior classification performance compared to Algorithm 2, particularly in minimizing false positives while maximizing detection of newly identified risky taxpayers.
• By reducing the number of reported taxpayers from 200 to 50, the method minimizes false positives, with 40 of the 50 correctly identified as entering the official risk list by December 2017.
• Reporting 2,000 taxpayers for audit using the method could identify up to 140 high-risk cases, demonstrating strong scalability and predictive power.
• The method enables automated, real-time fraud detection at the time of VAT return submission, supporting proactive auditing.
• The results suggest that network-based features significantly enhance fraud detection, highlighting the value of structural patterns in transaction networks.