As an emerging architecture, graph Transformers (GTs) have demonstrated significant potential in various graph-related tasks. Existing GTs are mainly oriented to graph-level tasks and have proved their advantages, but they do not perform well in node classification tasks. This mainly comes from two aspects: (1) The global attention mechanism causes the computational complexity to grow quadratically with the number of nodes, resulting in substantial resource demands, especially on large-scale graphs; (2) a large number of long-distance irrelevant nodes disperse the attention weights and weaken the focus on local neighborhoods. To address these issues, we proposed a new model, dual-branch graph Transformer (DCAFormer). The model divided the graph into clusters with the same number of nodes by a graph partitioning algorithm to reduce the number of input nodes. Subsequently, the original graph was processed by graph neural network (GNN) to obtain outputs containing structural information. Next, we adopted a dual-branch architecture: The local branch (intracluster Transformer) captured local information within each cluster, reducing the impact of long-distance irrelevant nodes on attention; the global branch (intercluster Transformer) captured global interactions across clusters. Meanwhile, we designed a hybrid feature mechanism that integrated original features with GNN outputs and separately optimized the construction of the query ( $Q$), key ( $K$), and value ( $V$) matrices of the intracluster and intercluster Transformers in order to adapt to the different modeling requirements of two branches. We conducted extensive experiments on 8 benchmark node classification datasets, and the results showed that DCAFormer outperformed existing GTs and mainstream GNNs.

Graph contrastive learning methods have recently emerged as a promising solution to tackle the problem of label scarcity in real-world scenarios. However, most of the existing methods are still flawed due to the lack of guided objectives. Moreover, they fail to effectively utilize complementary structural information from different graphs. To address these limitations, we propose a novel semi-supervised teaching graph contrastive network (STGCN) for node classification. Based on the teaching network architecture, STGCN establishes multi-level contrastive objectives, ensuring rich and detailed supervision for graph encoders. Specifically, after carefully analyzing the intrinsic correlation between different augmented views, we send a diffusion graph and two augmented views together into the novel teaching network, which owns one teacher encoder to guide two shared student encoders. Furthermore, we introduce a random sampling mixing module that extracts complementary information from multiple graphs, along with a label propagation technique to fully exploit limited labeled data. Finally, our method incorporates supervised contrastive loss and node similarity regularization to ensure coherent alignment between labeled and unlabeled nodes. Extensive experiments on five real-world node classification datasets demonstrate a maximum of 2.60% higher improvement than other models.