Quantum-Inspired Learning Detects Sparse-Ring Fraud in Transactions

Detecting coordinated fraud in financial transactions is challenging because individual transactions may appear benign, but a pattern emerges over time and across multiple entities. This study focuses on "sparse-ring fraud," a specific pattern where a completed directed cycle of transactions is distributed across several days, requiring models to integrate evidence from both temporal sequences and graph structures. Researchers developed an exploratory benchmark and a quantum-inspired Contextual Machine Learning (CML) prototype to address this. Using a synthetic transaction simulator that injects these sparse-ring patterns and includes broken-ring decoys, daily transaction graphs were aggregated into rolling windows. These were represented using raw graph features, persistent-homology summaries, or hybrid feature vectors. The study compared a Gated Recurrent Unit (GRU) baseline with the quantum-inspired CML as sequence-level classifiers. The exploratory results, based on synthetic data and a modest sample size, suggest that topology-only summaries are insufficient on their own, largely because they lose critical information like account-pair identity and edge direction. The most promising outcomes came from hybrid representations that combine identity-preserving graph features with topological summaries. This indicates that topological information is most effective when used as a contextual layer over dynamic graph features, and that CML is a promising model for fraud patterns whose evidence is distributed across temporal and relational contexts.