3C-GNN: Three-channel contrastive graph neural network for semi-supervised node classification

Abstract

In recent years, Graph Neural Networks (GNNs) have demonstrated exceptional performance in semi-supervised node classification. However, their effectiveness remains constrained by challenges such as insufficient supervision and representation collapse. Many methods solely focus on specific perspectives or single channels, failing to fully leverage the diversity of node features and topological structures. This results in incomplete learning of neighborhood information, limiting the expressive power of models. To address these issues, we propose a novel method called Three Channel Graph Contrastive Network (3C-GCN), characterized by three unique aspects: (i) A three-channel approach ,which captures more comprehensive information, thereby mitigating the risk of representation collapse and enhancing the model’s expressive capacity. (ii) The integration of contrastive learning with the three-channel method leverages self-supervised contrastive learning, enabling the model to effectively learn meaningful node representations without the need for extensive labeled data. This significantly improves performance on tasks with limited labeled samples. (iii) The pretraining-to-downstream task adaptation strategy enhances the model’s transfer learning capability, enabling it to generalize more effectively across various graph-related tasks, thereby strengthening its robustness and flexibility. Extensive experiments on multiple datasets demonstrate the effectiveness and generality of 3C-GCN compared to state-of-the-art methods.