Graph Neural Networks

Debates on oversmoothing, oversquashing, homophily/heterophily, and long-range tasks are muddled within each topic (e.g., oversquashing actually covers two distinct issues) so the paper urges the community to tease apart and precisely define these sub-problems.

Theoretical and empirical framework to analyze and perform graph rewiring in a principled way. Also, proposal of calulation of Commute Times (resistance) in a GNN layer and Bottleneck minimizarion using Spectral gradients.