FinFraudBench: A Heterogeneous Graph Benchmark for Financial Fraud Detection
The introduction of FinFraudBench, a new benchmark dataset, aims to overcome the shortcomings of current graph-based financial fraud detection standards. This dataset, elaborated in an arXiv paper (2608.15177), is crafted to authentically represent real-world financial systems by maintaining the complexity of multi-entity and multi-relational financial data. Unlike earlier benchmarks that often reduce financial ecosystems to simple, single-node graphs, FinFraudBench offers a comprehensive heterogeneous graph dataset that simulates realistic conditions, including severe class imbalance and scarce label availability. The authors highlight that existing public benchmarks do not accurately reflect real-world systems, lacking in both multi-entity preservation and large-scale heterogeneous datasets. FinFraudBench seeks to address these issues, providing a more precise platform for testing and developing graph-based fraud detection techniques.
Key facts
- FinFraudBench is a new heterogeneous graph benchmark for financial fraud detection.
- It is introduced in an arXiv paper with ID 2608.15177.
- The benchmark addresses limitations of existing graph-based fraud detection benchmarks.
- It preserves the multi-entity and multi-relational nature of financial data.
- It includes realistic operating conditions such as extreme class imbalance and limited label availability.
- The shift in fraud detection is from isolated transaction classification to relational risk reasoning.
- Graph-based models exploit dependencies among customers, cards, merchants, categories, and locations.
- Existing benchmarks often simplify financial ecosystems into homogeneous or single-node-type multi-relational graphs.
Entities
Institutions
- arXiv