CoDCL: Counterfactual-Inspired Augmentation Contrastive Learning for Temporal Link Prediction in Social Networks

Researchers have introduced CoDCL, a novel dynamic network learning framework designed to enhance temporal link prediction in rapidly evolving social networks. Current methods often fail to account for the causal mechanisms driving link formation, limiting their adaptability to complex, changing structures. CoDCL addresses this by integrating counterfactual-inspired data augmentation with contrastive learning. The framework employs a comprehensive strategy that combines dynamic treatments design with efficient structural neighborhood exploration to generate high-quality counterfactual data, effectively quantifying temporal changes in interaction patterns. A key feature of CoDCL is its plug-and-play architecture, allowing seamless integration into existing temporal graph models without requiring structural modifications. Extensive experiments on multiple real-world datasets demonstrate that CoDCL significantly outperforms state-of-the-art baselines. This advancement highlights the effectiveness of incorporating counterfactual reasoning into dynamic representation learning, offering a robust solution for predicting links in complex temporal environments where traditional algorithms struggle with emerging structural changes.