Towards Time-Variant-Aware Link Prediction in Dynamic Graph Through Self-supervised Learning

Guangqi Wen, Peng Cao, Zhiyong Jin, Ruoxian Song, Xiaoli Liu, Jinzhu Yang, Osmar R. Zaïane

Abstract: Dynamic graph link prediction is a challenging problem because the graph topology and node attributes vary at different times. A purely supervised learning scheme for the dynamic graph data usually leads to poor generalization due to insufficient supervision. As a promising solution, self-supervised learning can be introduced to dynamic graph analysis tasks. However, the self-supervised learning paradigm for dynamic graph learning has not been sufficiently investigated due to the complicated properties of the evolving graphs. We assume that the dynamic graphs consist of three independent types of key factors, i.e., graph time-variant information, time-invariant information and noise. Based on this assumption, we propose the Self-supervised Decoupling for Dynamic Graph (SDDG) framework for explicitly decoupling the representation which characterizes these three factors, thus enhancing the interpretability of the learned representation and link prediction performance. More specifically, we design an encoder-decoder architecture to sufficiently exploit the dynamic graph itself with multiple regularizations, so that the time-variant embedding of dynamic graph data can be effectively decoupled from the perspectives of node and structure for time-variant-aware link prediction. Experiments on five benchmark link prediction tasks demonstrate the significant improvement of our SDDG over the state-of-the-art methods. For example, SDDG achieves 98.2% and 97.3% Top-1 AUC on Reddit-B and Reddit-T, outperforming the DDGCL model, by 1.9% and 1.8%, respectively.