Renormalization Group as a principled way to avoid Oversquashing

Graph Neural Networks have proven to struggle to model long-range interactions. In this scenario they are forced to "over-squash" information into fixed-size vectors. We borrow a newly proposed Renormalization Group for graphs in order to shorten the distances of the interactions. We as well argue that this procedure allows us to separate the information related to different scales of the graph.

Join at: imt.lu/guinigi