Jacobi-enhanced dynamic hypergraph learning for graph-based fraud detection

Financial and e-commerce fraud detection remains challenging because fraudulent entities often exploit structural camouflage, coordinated behaviors, and complex relational dependencies. Existing graph-based approaches mainly rely on pairwise message passing, which may fail to explicitly characterize hidden higher-order affiliations and can suffer from unreliable information propagation under heterophilous graph structures. We propose JK-DHGNN, a dual-view graph learning framework designed for fraud detection scenarios where both observed structural relations and latent higher-order interactions are important. A Top- K dynamic hypergraph generator learns sparse latent affiliations, a Jacobi spectral filter aggregates multi-order graph signals, and a Jacobi polynomial-based KAN-like classifier provides adaptive nonlinear prediction. Experiments on YelpChi, Amazon, and Elliptic demonstrate that JK-DHGNN achieves competitive performance against representative graph-based fraud detection methods. Extensive analyses, including component ablation, dynamic structure analysis, perturbation evaluation, temporal case study, and computational-cost analysis, further reveal that the effectiveness of different modules depends on dataset characteristics and relational structures. These results highlight the potential of combining observed topology, latent higher-order modeling, and adaptive nonlinear classification for graph-based fraud detection.