SGCN: Structure and Similarity-Driven Graph Convolutional Network for Semi-Supervised Classification

Abstract

Traditional Graph Convolutional Networks (GCNs) primarily utilize graph structural information for information aggregation, often neglecting node attribute information. This approach can distort node similarity, resulting in ineffective node feature representations and reduced performance in semi-supervised node classification tasks. To address these issues, this study introduces a similarity measure based on the Minkowski distance to better capture the proximity of node features. Building on this, SGCN, a novel graph convolutional network, is proposed, which integrates this similarity information with conventional graph structural information. To validate the effectiveness of SGCN in learning node feature representations, two classification models based on SGCN are introduced: SGCN-GCN and SGCN-SGCN. The performance of these models is evaluated on semi-supervised node classification tasks using three benchmark datasets: Cora, Citeseer, and Pubmed. Experimental results demonstrate that the proposed models significantly outperform the standard GCN model in terms of classification accuracy, highlighting the superiority of SGCN in node feature representation learning. Additionally, the impact of different distance metrics and fusion factors on the models’ classification capabilities is investigated, offering deeper insights into their performance characteristics. The code and datasets are available at https://github.com/YONGLONGHU/SGCN.git.