Knowledge graphs (KGs) power downstream tasks such as question answering and recommender systems, but they are often very incomplete. To perform zero-shot knowledge graph completion in unseen KGs with different relational vocabularies, knowledge graph foundation models (KGFMs) have gained wide attention. A new paper on arXiv by Liu, Yinan; Xu, Wenjin; Zha, Zhiyuan; Yang, Xiaochun; and Wang, Bin introduces a simple yet effective adaptive negative sampling approach called KMAS to boost existing KGFMs.
The Challenge of Negative Sampling in KGFM Training
Existing KGFMs typically train using random negative triples, constructed by replacing the head or tail entity of a positive triple with a random entity. According to the paper, these negative triples are often of limited quality, providing weak supervision for KGFM training. This undermines the model's ability to generalize to new knowledge graphs. KMAS addresses this by constructing hard negative triples through updated relation embeddings generated from the existing KGFM's relation encoder.
How KMAS Adapts During Training
KMAS dynamically adjusts the ratio of hard negative triples throughout the entire training process. After a warmup phase, it increases the ratio linearly and then decreases it linearly. This adaptivity aligns with the evolving capability of the KGFM, ensuring that the model is challenged with appropriately difficult negatives at each stage.
Experimental Validation Across 44 Datasets
Extensive experiments were conducted over 44 datasets. The results demonstrate that KMAS can enhance many state-of-the-art (SOTA) KGFMs without requiring excessive additional time or memory consumption. The method is designed to be lightweight and plug-and-play, making it practical for enterprise deployments where computational resources are a concern.
Implications for Enterprise AI
While the paper focuses on academic benchmarks, improved knowledge graph completion has direct relevance for enterprise applications. KGFMs used in recommender systems, customer support, and internal knowledge bases can benefit from more accurate completion. The zero-shot capability is particularly valuable for organizations that need to integrate new domains without retraining from scratch. As enterprises increasingly rely on structured knowledge for decision-making, techniques like KMAS that improve model quality with minimal overhead are likely to see adoption in production pipelines.
| Feature | Random Negative Sampling | KMAS Adaptive Negative Sampling |
|---|---|---|
| Negative triple source | Random entity replacement | Hard negatives from relation embeddings |
| Adaptivity | Static | Dynamic ratio adjustment |
| Time/memory overhead | Baseline | No excessive additional consumption |
| Supervision quality | Weak (easy negatives) | Strong (hard negatives) |
The paper's authors are affiliated with institutions not specified in the source, but the work has been made available on arXiv under Computer Science > Artificial Intelligence. The code and data associated with the article are linked through the arXiv platform, enabling practitioners to reproduce and integrate the method into their own KGFMs.
For technology leaders evaluating AI infrastructure, the key takeaway is that improved negative sampling—a seemingly small component of the training pipeline—can yield significant gains in model performance for knowledge graph tasks. As KGFMs become a standard building block for enterprise AI, adopting methods like KMAS could reduce the cost of maintaining and updating knowledge bases while improving downstream task accuracy.