Boosting Knowledge Graph Foundation Models via Enhanced Negative Sampling

Researchers propose KMAS, an adaptive negative sampling method that enhances knowledge graph foundation models by constructing higher-quality hard negative triples and dynamically adjusting their ratio throughout training. The approach improves multiple state-of-the-art KGFMs across 44 datasets without significant computational overhead, advancing zero-shot knowledge graph completion for unseen relational vocabularies.

Knowledge graphs form the infrastructure for critical AI applications including question-answering systems and recommendation engines, yet their inherent incompleteness limits their utility. Foundation models trained on knowledge graphs face a significant challenge: generalizing to entirely new knowledge graphs with different relational vocabularies than those encountered during pre-training. Current approaches rely on random negative sampling during training, which produces weak supervisory signals that fail to effectively guide model learning.

The KMAS method addresses this limitation through an elegant two-part strategy. Rather than random negative sampling, it generates hard negative triples using updated relation embeddings from the existing model, forcing the system to learn more discriminative representations. The innovation extends beyond static hard negatives—KMAS implements an adaptive scheduling mechanism that increases hard negative ratios linearly after a warmup phase, then decreases them, aligning with the model's evolving learning capacity.

This research carries implications for the broader AI infrastructure landscape. Knowledge graph completion underpins enterprise search, semantic web applications, and retrieval-augmented generation systems increasingly important to modern AI workflows. By improving foundation model robustness without incurring substantial computational costs, KMAS enables more efficient scaling of knowledge-intensive AI systems. The comprehensive evaluation across 44 datasets demonstrates broad applicability rather than benchmark-specific optimization.

The work signals progress toward more practical foundation models that generalize across diverse knowledge domains without retraining. Future development should focus on combining adaptive negative sampling with other training improvements and testing against emerging larger knowledge graphs in production environments.

• →KMAS enhances knowledge graph foundation models through hard negative triple construction from updated relation embeddings rather than random sampling.
• →Adaptive scheduling dynamically adjusts hard negative ratios during training to match the model's evolving learning capability.
• →The method improves performance across 44 datasets without requiring additional time or memory overhead.
• →Knowledge graph foundation models now better generalize to unseen knowledge graphs with different relational vocabularies.
• →This advancement strengthens infrastructure for AI applications including question-answering, recommendations, and retrieval-augmented generation.