NodeBridge: Enhancing Semi-Supervised Learning on Graphs via Inter-node Learning

In recent years, graph neural networks have emerged as a prominent framework for semi-supervised learning of graph-structured data. However, conventional GNN models heavily depend on labeled data throughout the learning process while neglecting or minimally considering the available unlabeled data. To fully exploit this underutilized resource, this paper introduces NodeBridge, a framework designed to enhance graph semi-supervised learning and improve the performance of GNNs through inter-node learning. NodeBridge elevates GNN performance through graph data augmentation and contrastive learning. Graph data augmentation diminishes reliance on neighboring nodes by randomly discarding some nodes, whereas contrastive learning facilitates comprehensive information exchange between nodes via positive and negative sample pairs. NodeBridge effectively utilizes unlabeled data to enable nodes to acquire a broader spectrum of information, thus enhancing generalization. Additionally, it allows for direct connections between nodes through multi-hop random propagation and meticulously crafted contrast pairs. Experimental results demonstrate that NodeBridge achieves state-of-the-art performance in semi-supervised learning on GNN benchmark datasets.