Inductive link prediction on temporal networks through causal inference

The aim of inductive temporal link prediction is to forecast future edges associated with nodes unseen during training, which is a crucial task in the field of temporal network analysis. Existing methods mainly make predictions by learning from the node/edge attributes or investigating the node substructures. However, the deficiency of attributes limits the application scope of the attribute-aware methods, while the performance of the substructure-aware models is hindered by neglecting the node correlations or introducing structure bias. In addition, current inductive temporal link prediction methods struggle to generalize the learned network evolution pattern across different networks. To address these issues, we propose a Causal LInk Prediction (CLIP) framework for the inductive temporal link prediction task. Specifically, building upon the existing anonymous distance encoding strategy, we propose to eliminate the structure bias for estimating the true distance between nodes, which is achieved by the backdoor adjustment through the do-calculus, followed by decoupling the distance encoding vector to approximate the result. Moreover, to better adapt to the realistic scenarios, we further leverage the node substructures by considering the substructure features as the intervention on the basis of the true distance between nodes. In addition, our proposed approach achieves true inductive temporal link prediction by learning the universal evolution pattern across various temporal networks, which is accomplished through training on the synthetic dynamic graph data generated from the powerlaw cluster networks. We conduct extensive experiments on four real-world temporal networks, i.e., SuperUser, MathOverflow, AskUbuntu and StackOverflow, and the experimental results demonstrate that CLIP outperforms the baselines in terms of AP and AUC. In addition, experiments on the synthetic test graph data with various distributions showcase the remarkable generalization ability of CLIP.