.png){kind=icon}
Return
IF-NS: A New negative sampling framework for knowledge graph embedding using influence function
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.113258.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge graphs (KGs) are widely used in tasks such as recommendation and question answering, with the aid of knowledge graph embedding (KGE) models. These models incorporate negative sampling during training but face challenges, including vanishing gradients and false negatives, primarily due to the lack of incorrect facts in KGs. While score functions in KGE models help mitigate vanishing gradient issues, they often fail to address false-negative samples. Drawing inspiration from influence functions, we propose a new framework for translational distance KGE models called IF-NS, which leverages influence to select high-quality negative samples. The core idea is to assess how each sample contributes to the KGE model during training, effectively retaining high-quality negative samples. To reduce computational complexity, we also develop a simplified influence estimation algorithm tailored to the L2-norm distance in the KGE loss function. Furthermore, we introduce a repository for storing and tracking high-quality negative samples, employing repository-based sampling and dynamic updating strategies to balance exploration and exploitation. Experimental results show that IF-NS effectively mitigates the issues of vanishing gradients and false negatives, significantly improving KGE model performance.
Keywords:
Knowledge graph embedding
Negative sampling
Influence function
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSaveShareSaveShareSaveSwift and Sure: Hardness-aware Contrastive Learning for Low-dimensional Knowledge Graph Embeddings
OAAI
OAAIShareSave
View more