Revisiting Dynamic Graphs from the Perspective of Time Series

Weili Shi, Dongliang Guo, Sheng Li

Keywords: graph learning, dynamic graphs, time series

Abstract: Numerous studies have been conducted to investigate the temporal pattern of dynamic graphs. Existing methods predominantly fall into two categories: discrete-time dynamic graph (DTDG) methods and continuous-time dynamic graph (CTDG) methods. While these approaches have proven effective in modeling temporal dependencies within dynamic graphs, they exhibit several limitations. For instance, DTDG approaches often lose fine-grained temporal information. CTDG methods can preserve temporal details but may inadequately capture long-term dependencies due to computational constraints. Moreover, both paradigms predominantly focus on existing historical interactions, often neglecting the informative value of non-existing ones. These negative historical interactions can provide complementary insights into the recurring patterns of node behavior. To fully leverage both types of interactions, we propose transforming node interactions into binary time series. Building upon this formulation, we propose a novel model termed the Time Series-based Dynamic Graph (TSDyG) model, which approaches dynamic graph learning from a time series perspective. Compared to existing DTDG and CTDG methods, our model offers several advantages: it captures long-range dependencies, preserves fine-grained temporal details, and leverages information from both existing and non-existing historical interactions. We conduct extensive evaluations of our method on various benchmark datasets. The results demonstrate that our proposed TSDyG model achieves competitive performance on the downstream task such as link prediction.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 10022