Graph Navier-Stokes Networks

This work reformulates graph neural network message passing through the lens of the Navier-Stokes equations, decomposing propagation into convection for directed message transport and diffusion for local feature consistency. The model introduces node-level velocity fields on graphs and provides principled guidance for adapting GNN propagation across homophilic and heterophilic settings.

Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 6772-6781