Research on Link Prediction Based on Improved Graph Convolutional Network

Abstract

Link prediction is a fundamental task in network analysis with extensive applications in social networks, recommendation systems, and biological networks. In this work, we propose an improved Graph Convolutional Network (IGCN) model that leverages an encoder–decoder architecture, integrating multi-layer GCN-based aggregation of node features and local structure with adaptive hard negative sampling and contrastive learning. Our framework is evaluated on three benchmark datasets—Cora, Citeseer, and Pubmed—and achieves state-of-the-art performance when compared with traditional methods and other graph neural network models. This study demonstrates that incorporating adaptive negative sampling and contrastive loss effectively enhances the discriminative power of node representations for link prediction. Future research will focus on model scalability and the integration of domain-specific knowledge to further broaden its application scope.