CCEC-TGN for Link Prediction in Large-Scale Continuous-Time Dynamic Graphs

Abstract

Temporal link prediction aims to forecast future connections from historical interactions. As a fundamental problem in temporal graph learning, existing methods exhibit three major limitations. (1) They usually focus on node-level interaction histories and information aggregated from temporal neighbors, while overlooking cohesive information from node-associated subgraphs. (2) They do not sufficiently integrate high-order topological information with temporal information. (3) They are susceptible to confounding factors, which may lead to spurious correlations. To address these limitations, we propose CCEC-TGN, a temporal graph neural network designed for continuous-time dynamic graphs (CTDGs). (1) We leverage continuous coreness to capture time-evolving cohesive information. (2) We use structural encoding, conditional structural interpolation, and gated residual fusion to integrate macro-level structural information with micro-level temporal information. (3) We apply counterfactual consistency regularization to mitigate confounding factors and improve robustness and generalization. Experiments on several real-world temporal graph datasets demonstrate the effectiveness and efficiency of CCEC-TGN, especially in inductive prediction settings. The source code is available at https://github.com/BJHanjz/CCEC-TGN.