Self-Supervised Continuous Dynamic Graph Representation Learning via Hawkes Processes

Dynamic graph representation learning (DGRL) has garnered significant attention due to its prevalence in real-world applications. However, existing methods often rely on labeled data for training, which can be costly. Furthermore, these methods treat node links as instantaneous events, which limits their ability to capture node correlations and link persistence, resulting in suboptimal temporal representations in dynamic graphs. In real-world scenarios, the cumulative impact of past events on future outcomes can be superadditive, subadditive, or even subtractive, and can depend on the sequence in which events occur. Thus, we propose self-supervised DGRL (SDGRL), a novel self-supervised framework for DGRL based on Hawkes processes. SDGRL leverages Hawkes processes to model the interactions between events from multiple perspectives, resulting in more precise temporal representations. The key advantages of SDGRL include: it does not require labeled data for inductive representation learning; it captures the temporal interactions of events, including excitation and inhibition, through Hawkes processes; it learns versatile dynamic node embeddings that are not task specific but can generalize to various downstream tasks; and it is efficient, eliminating the need for complex encoders and significantly reducing memory and time costs. In addition, we demonstrate the stationarity and convergence of SDGRL, providing a theoretical foundation for our method. Extensive experiments on eight real datasets validate that the proposed method consistently achieves optimal performance compared to nine methods. Notably, on the Tmall dataset, SDGRL outperforms the state-of-the-art method, DyGFormer, by 28.73% in inductive dynamic link prediction accuracy while reducing time costs by a factor of 1.29 and memory usage by a factor of 15.33.