BETTER GNN GENERALIZATION

Summary from the TLDR AI newsletter

While maximizing spectral gap has been used to address over-squashing in GNNs, this paper shows that minimizing it can also improve generalization. The authors propose three new rewiring strategies to enhance label-community alignment, homophily, and computational efficiency.