Local-Global Geometric Insights for Graph Neural Networks via Entropic Curvature
The paper proposes entropic curvature as a global graph measure for how GNN information moves over distance.
Caich and Abbahaddou argue that common graph curvature tools stay too local to certify long-range propagation. They define a weak proxy meant to lower-bound global entropic curvature, then use it to derive bounds tied to oversmoothing, generalization, and graph expansion limits. The work turns the theory into three mechanisms: E-Gate aggregation, ENT structural encoding, and Midpoint-Completion Rewiring. It reports benchmarks against several rewiring and curvature methods on node- and graph-classification tasks. ArXiv · AI/CL/LG's note
Caich and Abbahaddou argue that common graph curvature tools stay too local to certify long-range propagation. They define a weak proxy meant to lower-bound global entropic curvature, then use it to derive bounds tied to oversmoothing, generalization, and graph expansion limits. The work turns the theory into three mechanisms: E-Gate aggregation, ENT structural encoding, and Midpoint-Completion Rewiring. It reports benchmarks against several rewiring and curvature methods on node- and graph-classification tasks. ArXiv · AI/CL/LG's note
score 4