Reproduce “Learning to Sample and Aggregate: Few-shot Reasoning over Temporal Knowledge 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
Ruijie Wang, Zheng Li, Dachun Sun, Shengzhong Liu, Jinning Li, Bing Yin, Tarek Abdelzaher
MetaTKGR introduces a framework for predicting future facts for newly emerging entities within Temporal Knowledge Graphs, learning from extremely limited initial observations. The model achieves an 11.4% relative gain over state-of-the-art baselines and demonstrates robust generalization over extended future periods by explicitly addressing data scarcity and dynamic distributional shifts.
Key insights
- Effectively addressing few-shot temporal reasoning requires simultaneously tackling data scarcity and dynamic distributional shifts over time.
- Meta-learning can be extended to dynamic graphs by explicitly modeling temporal adaptation through autoregressive optimization across time intervals and leveraging theoretical bounds like PAC-Bayes for stable generalization.
- Strategic temporal neighborhood sampling combined with time-aware attentive aggregation is crucial for learning robust representations of new entities from sparse, evolving data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.