KREPE: Generative Fact Generation for Hyper-relational Knowledge Graphs
A new research paper introduces KREPE, the first generative representation learning method for hyper-relational knowledge graphs (HKGs). Unlike traditional link prediction that assumes a single missing component, KREPE addresses the more realistic task of fact generation, where multiple or all components of a fact may be missing. The method uses masked discrete diffusion to model probability distributions of missing components conditioned on local and global graph structure. The paper is available on arXiv under ID 2605.24064.
Key facts
- KREPE is the first generative representation learning method for HKGs.
- The task is called fact generation: generating valid hyper-relational facts from arbitrarily masked queries.
- Traditional link prediction assumes nearly all entities and relations are known.
- KREPE uses masked discrete diffusion to model missing components.
- The method conditions on local fact components and global structure of HKGs.
- The paper is published on arXiv with ID 2605.24064.
Entities
Institutions
- arXiv