Aggregation-diffusion equations on graphs can exhibit metastable behavior, leading to clustered node representations that mitigate over-smoothing in graph neural networks.
