Physics-Informed Spatiotemporal Graph Neural Networks for Early Detection of Financial Fraud in Multi-Campus Shared Service Centers

The detection of financial statement fraud in multi-campus shared service centers presents unique challenges due to the intricate web of cross-campus fund flows, shared accounts, and reimbursement relationships that create complex interdependencies among organizational units. This study proposes a novel early warning framework that integrates Spatiotemporal Graph Neural Networks (STGNN) with Physics-Informed Neural Ordinary Differential Equations (PINN-ODE) to address these challenges. Our approach constructs dynamic heterogeneous graphs from monthly fund transfers, shared account transactions, and cross-campus reimbursements, with node features encompassing twelve-dimensional financial indicators including debt-to-asset ratios, budget execution deviations, and cash balance volatility. The STGNN module extracts spatiotemporal representations of anomalous signal propagation through graph convolution and temporal convolution operations, while the PINN-ODE solver embeds fundamental accounting conservation laws as differential equation constraints to ensure that state trajectories adhere to financial logic. The framework employs end-to-end joint optimization, utilizing trajectory deviation from the compliance manifold as a risk metric, combined with graph attention normalization to generate interpretable 0-1 fraud risk scores. Comprehensive experiments conducted on desensitized financial data from a large multi-campus university demonstrate that the proposed method achieves average Hit Rate@5 and Precision@5 values of no less than 0.90 and 0.89, respectively, across varying cross-campus correlation strengths. The early warning lead time for sudden fraud patterns averages approximately 3.8 days, demonstrating strong temporal responsiveness. These results establish the framework's effectiveness in enabling early, accurate, and interpretable detection of financial fraud across complex multi-campus organizational structures.