● PublishedSpringer · Book Chapter2026
Hybrid Spectral-Spatial Graph Neural Networks: Bridging Spectral Theory and Scalability
A. Soni, et al.
A Graph Neural Network that unifies spectral and spatial methods in one architecture — Laplacian eigenstructure for global graph geometry, localized message-passing for fine relational detail. It eases the usual expressiveness-vs-scalability trade-off, beating strong baselines on Cora, PubMed and ZINC with lower compute.
Earlier GNNs force a trade-off between expressiveness and scalability. Ours combines spectral graph theory (Laplacian eigenvalues and eigenvectors for global structure) with localized message-passing (first-order relational detail) in a single cohesive network. Benchmarked on Cora, PubMed and ZINC, it delivers consistent accuracy gains at lower computational overhead. Full details via the DOI above.