Reproduce “Event-Aware Prompt Learning for Dynamic Graphs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xingtong Yu, Ruijuan Liang, Renhe Jiang, Dongyuan Li, Yunxiao Zhao, Xinming Zhang, Yuan Fang

An Event-Aware Prompt Learning (EVP) framework is introduced for dynamic graphs, leveraging explicit historical event knowledge to bridge the objective gap between pre-training and diverse downstream tasks. It consistently outperforms state-of-the-art baselines on temporal link prediction and node classification across multiple datasets, achieving an AUC-ROC of 98.47% on Wikipedia transductive link prediction.

Key insights

  • Explicitly leveraging historical event knowledge through prompt learning significantly enhances dynamic graph model performance and task alignment, outperforming methods that rely on implicit event absorption.
  • Effective aggregation of historical events requires a dual consideration of temporal recency and underlying long-term patterns, as evidenced by the performance degradation when either the time decay or dynamic prompt is ablated.
  • The proposed framework offers parameter-efficient adaptability, allowing pre-trained dynamic graph encoders to efficiently align with diverse downstream tasks by tuning only a small set of prompt parameters.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.