.png){kind=icon}
Return
Gradient-decomposition based continual embedding method for growing knowledge graphs
.png){kind=icon}
PRE
AI
DOI：10.1007/s13042-026-02997-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Existing knowledge graph (KG) embedding models are typically designed to facilitate the representation learning of static KGs. However, KGs tend to evolve and expand over time, and existing methods are unable to adapt to the growing KGs. How to obtain useful information from new facts and inject them to the previous-learned embeddings is critical but challenging for the effective representation of growing KG. In this paper, we proposed a continual embedding method of knowledge graph to efficiently learn information from the new triples and overcome the catastrophic forgetting problem on the existing triples. By dividing the overall loss of the embedding model into the loss on the previous triple set and the newly-arrived triple set, we transform the continual embedding task into a constrained optimization problem. And then a deft working-set decomposition and gradient decomposition strategy is proposed to efficiently train the continual embedding model. Three benchmark datasets are used to evaluate the performance of the proposed method, and the results demonstrate that it exhibits advantageous performance compared to the state-of-the-art models.
Keywords:
Growing knowledge graphs
Representation learning
Embedding
Continual learning
Incremental learning
Journal
IF:
2.7
Papers:
3.2K
Citations:
5.6K
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSave
View more