Three-channel Residual Graph Convolutional Network for Semi-supervised Node Classification

Graph convolutional network (GCN) has recently gained significant success in node classification tasks. However, it still suffers from the problem of over-smoothing when stacking multiple layers, as well as the challenge of effectively integrating deeper-level relevant information between node features and the topological structure. To address these issues, this paper put forward a novel three-channel residual graph convolutional network (TR-GCN) model by removing the non-linear activation functions in deep GCN and adopting residual feature propagation, leading to the development of a residual graph convolutional structure. In the structure, a k-nearest neighbors feature graph is pre-built based on the similarities between nodes and combine it with the original topological graph as input to the parameter-shared residual graph convolutional module to extract common information. Additionally, an attention mechanism is introduced to adaptively adjust the weights of different embeddings, enabling comprehensive fusion of feature information. This attention mechanism allows the model to fully leverage information from both labeled and unlabeled nodes in semi-supervised node classification tasks, thereby improving classification accuracy. Extensive experiments are conducted on three benchmark datasets, and the results demonstrate that TR-GCN outperforms the baseline methods.