Enhancing link prediction in complex networks using GraphSAGE with graph diffusion convolution

Abstract

Link prediction plays a fundamental role in numerous applications, including social network analysis, recommendation systems, and dynamic interaction modeling, by enabling the analysis and prediction of the evolution of complex networks. Existing Graph Neural Network (GNN)-based methods typically fail to adequately model long-range structural dependencies while remaining tractable and reliable on sparse, changing graphs. To address these limitations, this paper proposes a diffusion-enhanced inductive link prediction framework that combines Graph Diffusion Convolution (GDC), structural node descriptors, and neighborhood aggregation from GraphSAGE. The framework first extracts multiple topological features, such as degree, clustering coefficient, PageRank, eigenvector centrality, betweenness centrality, and core number, and then computes Personalized PageRank. They used a diffusion preprocessing method to leverage higher-order structural connectivity without sacrificing sparsity through threshold-based pruning. The diffusion-enhanced graph is then passed to a multi-layer GraphSAGE encoder to learn inductive node representations, and the negative edge sampling and binary cross-entropy loss are used to optimize the link probabilities. This model achieves higher accuracy than the other models on the benchmark datasets. Experimental results show that the Proposed model outperforms other models on the benchmark datasets. It regularly improves the metrics of AUC, Precision, Recall, and F1-score by ~ 2.45% on average, over the best competing baseline, especially when used on networks with strong community structures. The results of additional ablation and sensitivity analyses show that diffusion-based structural propagation makes a significant contribution to the observed performance improvement.