🤖 machine learning

TED: Related Party Transaction guided Tax Evasion Detection on Heterogeneous Graph

This paper proposes TED, a novel graph neural network model that leverages heterogeneous graphs and a hierarchical attention mechanism to analyze related party transactions, significantly improving tax evasion detection performance by capturing complex interactive information that traditional statistical methods miss.

Original authors: Yiming Xu, Bin Shi, Bo Dong, Jiaxiang Wang, Hua Wei, Qinghua Zheng

Published 2026-05-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yiming Xu, Bin Shi, Bo Dong, Jiaxiang Wang, Hua Wei, Qinghua Zheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the global economy as a giant, bustling marketplace. In this marketplace, most people play by the rules, but some shopkeepers are trying to sneakily hide their profits to avoid paying their fair share of taxes. This is tax evasion, and it's like a thief stealing from the public treasury, leaving less money for schools, roads, and hospitals.

For a long time, tax authorities tried to catch these thieves by looking at individual shopkeepers' ledgers (their financial data) and asking experts to write down a list of "suspicious signs." If a shopkeeper had too many red flags, they got flagged. But this method was like trying to find a needle in a haystack by only looking at the needle's color, ignoring the fact that the needle might be hidden inside a specific type of box.

This paper introduces a new, smarter way to catch these tax evaders, called TED (Tax Evasion Detection). Here is how it works, explained through simple analogies:

1. The Problem: It's Not Just One Shopkeeper

The authors realized that tax evasion rarely happens in isolation. It usually involves a web of relationships. A "bad" shopkeeper might be connected to a "bad" supplier, a "bad" accountant, and a "bad" investor.

The Old Way: Looking at one shopkeeper in a vacuum.
The New Way (TED): Looking at the entire web of connections.

The authors built a giant map called a Heterogeneous Graph. Imagine this map not just as dots and lines, but as a complex city with different types of buildings:

Companies (The shops)
People (The owners and investors)
Items (The goods being sold)
Events (The transactions)

In this city, lines connect these buildings. A person owns a company; a company buys items; a company sells to another company.

2. The Challenge: Too Much Noise

The problem with this giant city map is that it's noisy. There are millions of innocent connections. A company might buy a pencil from another company just because it's a normal business deal. If you look at every connection, you get overwhelmed by the noise and miss the real criminals.

The paper identifies a specific type of connection that is the "smoking gun": Related Party Transactions (RPT).

Analogy: Imagine two shopkeepers who are secretly brothers-in-law. They sell goods to each other at fake prices to move money around and hide profits. To an outsider, it looks like a normal sale, but to the tax man, it's a suspicious "family reunion" of money.

3. The Solution: The "TED" Detective

The authors created a system called TED that acts like a super-smart detective. It doesn't just look at the whole city; it focuses on specific, suspicious neighborhoods called RPT Groups.

Here is how TED solves the two main problems:

A. Filtering the Noise (The "High-Order" Filter)

Most computer programs look at neighbors who are just one step away (e.g., "Who did Company A sell to?"). TED looks deeper. It looks for RPT Groups—specific patterns of connections that are known to be used for tax evasion (like a person owning two companies that trade with each other).

The Metaphor: Imagine you are looking for a thief. A normal search looks at everyone who walked past the thief. TED ignores the crowd and only looks at the thief's "inner circle" of accomplices. By focusing on these specific groups, TED filters out the innocent "noise" of normal business.

B. The Hierarchical Attention (The "Smart Magnifying Glass")

Once TED finds these suspicious groups, it needs to figure out which parts of the group are the most dangerous.

Inner-RPT Level: TED looks inside a single suspicious group. It asks, "In this group of five people, which two are doing the most suspicious trading?" It uses a multi-head attention mechanism (think of it as having multiple magnifying glasses looking at different angles) to weigh the importance of each connection.
Cross-RPT Level: TED then looks at the whole company. A company might be involved in five different suspicious groups. TED asks, "Which of these five groups is the biggest threat?" It assigns a "weight" or score to each group to decide how much it should influence the final verdict.

4. The Results: Catching More Thieves

The authors tested TED on real data from tax bureaus in China. They compared it against 13 other top-tier methods (like standard AI models and expert rule systems).

The Outcome: TED was significantly better at catching tax evaders. It improved the detection accuracy by about 7.6% and the F1 score (a measure of how well it balances catching thieves without falsely accusing innocent people) by 8.5% compared to the best existing methods.
Why it worked: Because it stopped looking at isolated data points and started understanding the complex, hidden relationships between people and companies.

Summary

Think of tax evasion detection as finding a needle in a haystack.

Old methods tried to find the needle by looking at the color of the straw.
TED realized the needle is usually hidden inside a specific, weirdly shaped box (the RPT group). It ignores the rest of the haystack, zooms in on those boxes, and uses a smart magnifying glass to see exactly which part of the box contains the needle.

The paper concludes that this method is so effective it has already been applied to real risk management systems in tax bureaus, proving that understanding the "who knows who" in the business world is the key to stopping tax fraud.

Technical Summary: TED: Related Party Transaction guided Tax Evasion Detection on Heterogeneous Graph

Problem Statement
Tax evasion causes significant government revenue losses and disrupts fair economic competition. While existing detection solutions rely on expert-designed statistical features or network-based methods, they often fail to fully exploit the rich, interactive information inherent in tax scenarios. Specifically, traditional methods struggle to model the complex, compound, and covert relationships between taxpayers, particularly within Related Party Transaction (RPT) groups. Furthermore, heterogeneous tax data contains substantial noise, where irrelevant relationships obscure the signals necessary for accurate detection. The paper addresses the challenge of detecting tax evasion by modeling tax scenarios as heterogeneous graphs and leveraging RPT groups to filter noise and capture deep structural semantics.

Methodology: The TED Framework
The authors propose TED (RPT-guided Heterogeneous Graph Hierarchical Attention for Tax Evasion Detection), a novel Graph Neural Network (GNN) model designed to operate on heterogeneous graphs containing multiple entity types (e.g., companies, persons, items, events) and relationship types. The model operates through two primary hierarchical levels:

RPT Group as High-Order Proximity:
Instead of relying on simple metapaths or $k$ -order neighbors, TED defines RPT groups (e.g., Person-Company-Company-Person) as high-order proximity structures. This approach filters out low-order noisy relationships that are irrelevant to tax evasion, focusing instead on complex transaction groups where tax evasion is more likely to occur.
Hierarchical Attention Mechanism:
To capture the compound influences within and across these RPT groups, TED employs a two-level attention framework:
- Inner-RPT Level: This level extracts features from individual RPT group instances surrounding a target node. It utilizes a subgraph matching approach to identify instances and employs a multi-head attention mechanism to weigh the importance of different nodes within a specific RPT instance. It aggregates features from all neighbors within the instance, preserving the heterogeneous topology rather than collapsing it into a homogeneous graph.
- Cross-RPT Level: This level aggregates information across different RPT groups. Since different RPT groups contain distinct topologies and semantic meanings, a cross-RPT attention mechanism is used to learn the relative importance of each RPT group type for the target node. It combines the original node features with the aggregated RPT features to compute attention coefficients, ultimately producing a final node embedding.

The model is trained using a semi-supervised learning approach, minimizing cross-entropy loss on labeled tax evasion data while utilizing the graph structure to propagate information to unlabeled nodes.

Key Contributions
The paper outlines three main contributions:

Heterogeneous Graph Modeling: The authors are the first to study tax evasion detection under a heterogeneous graph model incorporating over 10 kinds of nodes and relations, arguing that this comprehensive representation improves detection performance compared to feature-based or homogeneous graph methods.
RPT-Guided Hierarchical Attention: The proposal of using RPT groups as high-order proximity to filter low-order noise, coupled with a novel hierarchical attention model (Inner-RPT and Cross-RPT levels) to mine deep structural and semantic information from complex, covert relationships.
Real-World Application and Evaluation: The method was applied to a real risk management system of a Chinese tax bureau and evaluated on two human-labeled, real-world tax datasets (T20H and T15S).

Experimental Results
The authors evaluated TED against 13 state-of-the-art baselines, including proximity-preserving methods (e.g., HIN2Vec, Metapath2vec), relation-learning methods (e.g., TransE, DistMult), PU learning methods (e.g., PUNE), and message-passing GNNs (e.g., GCN, HAN, HGT).

Performance: TED significantly outperformed all baselines. On the T20H and T15S datasets, it achieved average F1 score improvements of 8.50% and 12.86%, and accuracy improvements of 7.61% and 11.40% over the best existing methods, respectively.
Robustness: The model demonstrated stability across varying positive sample ratios (PSR) in the training set, maintaining superior performance even when labeled data was scarce (e.g., 10% PSR).
Ablation Studies: Experiments confirmed the necessity of each component. Removing the heterogeneous node types (TED-hete), the attention mechanisms (TED-att), or specific attention levels (TED-inner, TED-cross) resulted in significant performance degradation.
Visualization: t-SNE visualizations showed that TED produced clearer separation between tax-evading and normal companies compared to other models, indicating superior representation learning.
Efficiency: TED demonstrated high efficiency, achieving comparable performance to other models in significantly less training time and showing linear scalability with graph size.

Significance and Claims
The paper claims that the primary significance of this work lies in shifting the paradigm of tax evasion detection from static feature extraction to dynamic, structure-aware modeling using heterogeneous graphs. By explicitly modeling RPT groups as high-order proximities, the method effectively addresses the "noisy relationship" and "compound relationship" challenges inherent in tax data. The authors assert that their approach not only improves detection accuracy but also provides a framework that can be integrated into real-world tax risk management systems. They conclude that the hierarchical attention mechanism successfully uncovers the hidden semantic information within complex tax networks, offering a robust solution for a global economic problem. Future work is suggested to explore reinforcement learning for the autonomous selection of meaningful RPT groups, potentially reducing the need for domain expert intervention.