Adaptive Negative Sampling Enhances Knowledge Graph Foundation Models
A new method called KMAS (Knowledge graph Model Adaptive negative Sampling) improves knowledge graph foundation models (KGFMs) by generating higher-quality negative triples during training. Knowledge graphs (KGs) are crucial for tasks like question answering and recommender systems but often suffer from incompleteness. KGFMs aim to perform zero-shot completion on unseen KGs with different relation vocabularies. Existing KGFMs typically use random negative triples, which provide weak supervision. KMAS constructs hard negative triples using updated relation embeddings from the KGFM's relation encoder, offering a simple yet effective enhancement. The paper is available on arXiv under reference 2605.27023.
Key facts
- KMAS is an adaptive negative sampling approach for KGFMs.
- Knowledge graphs are often incomplete.
- KGFMs perform zero-shot completion on unseen KGs.
- Random negative triples provide weak supervision.
- KMAS uses updated relation embeddings to construct hard negatives.
- The method is simple yet effective.
- The paper is on arXiv: 2605.27023.
- KGs support question answering and recommender systems.
Entities
Institutions
- arXiv