Mixture of Scope Experts at Test: Generalizing Deeper Graph Neural Networks with Shallow Variants

The study published on arXiv explores the limitations of graph neural networks (GNNs) when dealing with heterophilous graphs, where dissimilar nodes connect. It highlights that while increasing the depth of GNNs can potentially expand their receptive fields, this often results in performance degradation. The research indicates that deeper GNNs yield only marginal improvements compared to their shallower counterparts. To address these issues, the authors introduce a novel approach called Mixture of Scope Experts at Test (Moscat), which aims to enhance the generalization capabilities of deeper GNNs while maintaining their expressivity. Experimental results demonstrate that Moscat effectively improves accuracy across a diverse range of GNN architectures and datasets, showcasing its flexibility and potential for broader applications in the field of artificial intelligence.