Boosting Knowledge Graph Foundation Models via Enhanced Negative Sampling

The paper introduces KMAS, an adaptive negative sampling method designed to improve knowledge graph foundation models (KGFMs) by constructing hard negative triples using updated relation embeddings. This approach dynamically adjusts the ratio of hard negatives during training, leading to enhanced performance across 44 datasets without significant additional resource requirements. The findings are significant for practitioners as they provide a more effective training strategy for KGFMs, improving their ability to handle incomplete knowledge graphs in zero-shot scenarios.