Enhance Message Passing with Cluster-Aware Virtual Nodes for Semi-Supervised Node Classification

Graph Neural Networks (GNNs) excel at semi-supervised node classification but struggle with modeling long-range dependencies with message-passing mechanisms. Despite attempts to tackle this issue through deeper models or new architectures like graph transformers, they often introduce high computational complexity or provide unsatisfactory results. While a single virtual node connected to all other nodes is a simple and effective approach for graph classification, it is less suitable for node classification, as the latter task demands more fine-grained information and the effective utilization of both local and high-order connections. To address it, we propose CLuster-aware vIrtual Nodes Enhanced graph neural networks (CLINE), which uses multiple virtual nodes to distribute global information and mitigate the over-squashing problem. We further introduce a three-stage message-passing framework that leverages these cluster-aware virtual nodes to integrate both high-order and local information. Extensive experiments on various synthetic and real datasets demonstrate that CLINE significantly outperforms existing methods in semi-supervised node classification tasks1.