Interactions Exhibit Clustering Rhythm: A Prevalent Observation for Advancing Temporal Link Prediction

Siwei Zhang, Xi Chen, Yun Xiong, Xixi Wu, Yizhu Jiao, Yao Zhang, Mingyang Zhang, Tengfei LIU, Weiqiang Wang, Jiawei Zhang

Keywords: Link Prediction, Temporal Graphs, Data Mining

TL;DR: Based on the introduced empirical analyses, we observe a widespread and distinct temporal clustering in node interaction rhythms and leverage this interesting phenomenon for advancing temporal link prediction.

Abstract:

Temporal link prediction aims to forecast future link existence in temporal graphs, with numerous real-world applications. Existing methods often rely on designing complex model architectures to parameterize the interaction patterns between nodes. Instead, we re-think the interaction dynamics in temporal graphs (which we call ``interaction rhythms'') by addressing a fundamental research question: \textit{Is there a strong yet prevalent latent interaction rhythm pattern across different temporal graphs that can be leveraged for temporal link prediction?}
Our introduced empirical analyses reveal that there indeed exists temporal clustering in node interaction rhythms, where for a specific node, interactions tend to occur in bursts. Such observation leads to two key insights for predicting future links: (i) recent historical links that carry the latest rhythm pattern information; and (ii) the inter-event times that further illuminate temporal dynamics.
Building on these empirical findings, we propose TG-Mixer, a novel method that explicitly captures temporal clustering patterns to advance temporal link prediction.
TG-Mixer samples the most recent historical links to extract surrounding neighborhoods, preserving currently invaluable interaction rhythms while avoiding massive computations.
Additionally, it integrates a carefully designed silence decay mechanism that penalizes nodes' long-term inactivity, effectively incorporating temporal clustering information for future link prediction.
Both components ensure concise implementations, leading to a lightweight architecture.
Exhaustive experiments on seven benchmarks against nine baselines demonstrate that TG-Mixer achieves state-of-the-art performance with faster convergence, stronger generalization capabilities, and higher efficiency. The experimental results also highlight the importance of explicitly considering temporal clustering for temporal link prediction.

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5734