LCLRD: Link Prediction via Contrastive Learning With Relational Distillation

Link prediction is widely used in various fields to predict the missing or potential future links between node pairs in the network. Knowledge distillation (KD) methods have been introduced for the link prediction task, demonstrating exceptional performance in inference acceleration. However, these methods primarily focus on supervised learning settings that rely on high-quality labels, and they fail to capture the rich structural features inherent in graph data within the teacher model effectively, resulting in suboptimal performance. To address the problem of reducing label dependency, we propose a link prediction method via contrastive learning with relational distillation (LCLRD) that aims to improve model performance under label scarcity. LCLRD incorporates contrastive learning into the teacher model for pretraining, enabling the teacher model to obtain high-quality node representations more effectively. Then, graph structural information is extracted from the teacher model via relational distillation and transferred to the student model. In addition, we leverage the Pearson correlation coefficient (PCC) as a new matching strategy to replace the Kullback–Leibler (KL) divergence, aiming to alleviate the prediction discrepancy between the student and the stronger teacher model. The experimental results show that LCLRD significantly outperforms other baseline methods on 12 datasets, demonstrating the superiority of this approach.