Revisiting Dynamic Graphs from the Perspective of Time Series

TMLR Paper9012 Authors

Abstract: Numerous studies have investigated temporal modeling in dynamic graphs. Existing approaches predominantly fall into two categories: discrete-time dynamic graph (DTDG) methods and continuous-time dynamic graph (CTDG) methods. While both paradigms have shown effectiveness in capturing temporal dependencies, they suffer from several inherent limitations. Specifically, DTDG approaches often lose fine-grained temporal information due to snapshot-based discretizations, whereas CTDG methods preserve precise timestamps but may struggle to capture long-range temporal dependencies because of computational constraints. Moreover, interactions in real-world dynamic graphs frequently exhibit predictable and recurring temporal patterns, which are not fully exploited by existing methods. To better leverage such regularities, we propose to transform node interactions into binary time-series representations, enabling explicit modeling of temporal patterns. Building on this formulation, we introduce a novel model, termed Time Series-based Dynamic Graph (TSDyG), which approaches dynamic graph learning from a time-series perspective. Compared to existing DTDG and CTDG methods, TSDyG offers several advantages: it preserves fine-grained temporal information, captures long-range dependencies, and effectively capture recurring interaction patterns. We conduct extensive experiments on multiple benchmark datasets, and the results demonstrate that TSDyG achieves competitive performance on downstream tasks such as temporal link prediction.

Submission Type: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission: N/A

Assigned Action Editor: ~Mingming_Gong1

Submission Number: 9012