Enhancing Low-Degree Graph Neural Networks via Joint Training and Improved Message Passing

Abstract

Graph Neural Networks (GNNs) have achieved significant success in tasks involving graph-structured data, such as node classification and link prediction. However, low-degree graphs, which are common in real-world situations, pose a significant challenge to GNN performance by limiting information dissemination and hindering the capture of long-range dependencies. In this paper, we propose a novel method, JT-IMPN (Joint Training Integrated Message Passing Network), to address this issue. Firstly, we introduce IMPM (Improved Message Passing Mechanism), which improves the message-passing mechanism by incorporating residual connections and multi-hop embeddings, thereby generating more accurate node embeddings and mitigating the problem of over-smoothing. Secondly, we construct a global-connectivity graph utilising the node embeddings obtained from IMPM, to enhance the reliability of the constructed graph. Additionally, during the construction process, to enhance overall robustness, the three enhanced graphs were fused into a unified global-connectivity graph. Finally, we propose a joint training framework that integrates low-degree and global-connectivity graphs, benefiting from both local and global information while mitigating the impact of excessive information from neighboring nodes. The experimental results for the node classification task highlight the superiority of our model.