Dynamic Rescaling for Training GNNs

For heterophilic graphs, achieving balance based on relative gradients leads to faster training and better generalization. In contrast, homophilic graphs benefit from delaying the learning of later layers.