In this paper, the authors explore the use of Graph Neural Networks (GNNs) in making predictions on graph data. They specifically examine cases where the graph structure is not informative for the predictive task, such as in molecular properties. They find that GNNs tend to overfit the graph structure, even when ignoring it would lead to better results. The authors provide empirical evidence and a theoretical explanation for this phenomenon, and propose a graph-editing method to address it. This method improves the accuracy of GNNs in various benchmarks. Overall, this work highlights the challenges and potential solutions in using GNNs for predictions on graph data.
https://arxiv.org/abs/2309.04332