Ramanujan Graph Rewiring with Non Negative Resistance Curvature

This article presents Ramanujan Propagation, a novel graph rewiring method utilizing Ramanujan graphs to address the issue of over-squashing in Graph Neural Networks (GNNs). By ensuring non-negative resistance curvature, the proposed approach enhances long-range dependency learning and maintains local connectivity during graph rewiring. Experimental results show that this method surpasses nine existing rewiring techniques, highlighting its potential for improving scalable and topology-aware message passing in GNNs.