FinFraudBench: A Heterogeneous Graph Benchmark for Financial Fraud Detection
This paper introduces FinFraudBench, a new heterogeneous graph benchmark featuring two large-scale, realistic financial datasets with multi-entity types and directed edges to address the limitations of existing benchmarks in capturing the complexity, class imbalance, and label scarcity of real-world financial fraud detection.
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
In the vast, invisible network of modern finance, every swipe of a card or transfer of money leaves a digital footprint. For decades, experts have tried to catch fraudsters by looking at these footprints one by one, treating each transaction as an isolated event. They would examine the amount, the time, and the location, hoping to spot a pattern that didn't fit. But fraud is rarely a solitary act; it is a web of connections. A stolen credit card is used by a specific person, at a specific store, often in a specific way that links it to other suspicious activities. To understand the whole picture, researchers realized they needed to stop looking at single records and start mapping the relationships between the people, the cards, the merchants, and the places involved. This shift from checking individual items to understanding how they connect is the foundation of graph-based detection, a method that treats financial data not as a spreadsheet, but as a complex map of interactions.
A team of researchers has now built a new, more realistic map to test these detection methods, addressing a critical gap in how fraud is studied. They created a benchmark called FinFraudBench, which consists of two massive datasets constructed from real-world transaction records. Unlike previous tools that simplified the financial world into a single type of node or a single kind of relationship, these new datasets preserve the messy, multi-layered reality of finance. They contain nearly 9 million nodes representing different entities like customers, cards, merchants, categories, and locations, linked by nearly 90 million directed connections. The researchers carefully organized this data to mimic the difficult conditions of the real world, where fraud is extremely rare compared to legitimate transactions and where only a tiny fraction of cases are labeled as known fraud.
The researchers used this new benchmark to test a wide variety of existing detection models, ranging from simple statistical tools to advanced artificial intelligence systems designed to read graphs. They found that methods which ignored the different types of entities and treated everything as the same were significantly less effective. The most successful models were those that could distinguish between a customer, a card, and a merchant, understanding that each plays a different role in the network. Specifically, models designed to handle heterogeneous graphs—those that respect the different types of nodes and the specific ways they connect—consistently outperformed their simpler counterparts. These advanced models were better at ranking risky transactions and identifying fraud even when the data was heavily imbalanced, a common and difficult scenario in real banking systems.
The study suggests that the key to catching financial fraud lies in preserving the rich, typed structure of the data rather than flattening it. When researchers forced the complex, multi-entity data into a simpler format, they lost crucial signals that helped identify bad actors. The results indicate that the most promising path forward is to build systems that can simultaneously process the different types of financial entities and their unique relationships. While some specialized models designed to spot specific fraud patterns remained competitive, they generally could not match the overall performance of the models that fully embraced the complexity of the heterogeneous graph. This work provides a standardized way for the scientific community to evaluate future tools, ensuring that new methods are tested against a realistic representation of the financial ecosystem rather than an oversimplified version.
The researchers constructed these datasets from public transaction records, converting rows of data into a network where transactions are the central points connected to the entities involved. They ensured that no information leaked from the future into the past, a common pitfall in fraud detection, by strictly separating the data into training, validation, and testing sets. The final datasets included over 1.8 million transactions for one dataset and nearly 9 million for the other, with fraud rates as low as 0.15 percent, mirroring the extreme rarity of fraud in real life. By testing a broad spectrum of algorithms on this terrain, the team was able to show that the ability to navigate the different types of connections is not just a theoretical advantage but a practical necessity for effective fraud detection. The findings point toward a future where financial security systems are built on a deeper understanding of the diverse relationships that define the digital economy.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.