Unifying Graph Neural Networks Through a Common Layer Equation

arXiv:2608.160972026-08-16

Graph neural networks are commonly described through family-specific equations whose notation obscures shared computations and structural differences. We introduce a common layer equation that represents covered architectures through seven components: an update domain, channel set, propagation bank, per-channel message maps, channel-fusion operator, ego/residual map, and update map. The central factorization separates where information moves, encoded by the propagation bank, from what moves, enc

Authors · Sai Karthik Navuluru