Leveraging long-range nodes in multi-view graph contrastive learning
Ludan He, Debo Cheng, Guixian Zhang, Shichao Zhang
Abstract
Graph convolutional networks (GCNs) have achieved encouraging results in various graph learning tasks. However, their performance significantly decreases in scenarios requiring long-range interactions. For instance, in chemistry, molecular properties may depend on interactions between distant atoms, which conventional GCNs struggle to capture effectively. Most existing studies attempt to address this by constructing high-order views through multi-view learning. Yet, such modifications may disrupt the original graph data’s potential low-dimensional representation. To address these challenges, we have developed a novel method called Leveraging Long-range Nodes in multi-view graph Contrastive Learning (LLNCL). LLNCL achieves comprehensive and high-quality modeling of node features and structures. It facilitates long-range information transmission between nodes and dynamically updates the constructed graph to enhance its quality. Moreover, LLNCL preserves the original structural features, enabling the joint learning of both short-range and long-range information. Additionally, through multi-view contrastive learning, LLNCL dynamically updates the graph structure, effectively filtering out unreliable information to obtain higher-quality node features. Extensive experiments on six real-world datasets demonstrate that our method outperforms current state-of-the-art semi-supervised classification methods.