Scaling GNNs with Graph Rewiring

This article is aimed at machine learning engineers. In the most of classical deep learning; models showed an increase in performance with addition of more layers. In Graph neural networks however addition of more layers shows a substantial drop in performance compared to shallow models . Ideally,we would need more layers to model complex relationships, long-range information. When we add more layers to a GNN each node now has access to information from other nodes that are farther away and might help model important long-range dependencies.