Relationship updating network with contrastive learning

In Graph Neural Networks (GNNs), a common feature across many datasets is the Power-law Distribution of node degrees, where most nodes exhibit few connections, contrasting with a small fraction that possesses a high number of links. This difference often introduces training instability and compromises performance on tasks like node classification, particularly for low-degree nodes. To tackle these challenges, we introduce RUNCL: Relationship Updating Network with Contrastive Learning, a novel model designed to ensure that the model learns more accurate node features, especially the features of low-degree nodes. Specifically, RUNCL comprises a graph generation module that generates different neighborhood information graphs based on the node feature graphs. The optimal graph selection module selects the neighborhood information graph that best reflects the relationship between nodes and a contrastive learning module to learn more accurate node embeddings by contrasting positive and negative samples. We evaluate the performance of RUNCL on six datasets, and the experimental results demonstrate its effectiveness. The model exhibited an improvement of 2.5% in the testset. Moreover, the model’s performance boosted to 6% when the testset only included low-degree nodes. The implementation and data are made available at https://github.com/pengyu-zhang/RUNCL-Relationship-Updating-Network-with-Contrastive-Learning.