.png){kind=icon}
.png){kind=icon}
Return
An Efficient and Lightweight Framework for Unsupervised Entity Alignment between Temporal Knowledge Graphs
.png){kind=icon}
PRE
AI
DOI：10.1016/j.inffus.2025.103821.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Entity Alignment (EA) aims to build a content-rich unified knowledge graph (KG) by linking entities that have the same semantics but belong to different KGs, which is a foundational task in KGs fusion. Temporal knowledge graphs (TKGs) expand static KGs by adding timestamps. Existing methods have attempted to embed temporal information into EA and have made significant progress. Most existing methods achieve entity alignment by pre-aligning entity pairs, which often requires prior knowledge, significantly limiting the applicability of these methods. Emerging methods that generate pseudo-alignments also tend to generate wrong seeds, which affect training results. In addition, as the scale of TKGs continues to expand, the efficiency of EA models becomes more and more important. To address these issues, we propose an Efficient and Lightweight framework for unsupervised EA between TKGs (ELTEA). Specifically, to make full advantage of the structural information of TKGs, we employ relationship-aware graph attention networks to capture the interactions between entities and relationships. We then utilize Jaccard similarity to calculate the temporal similarity matrix. By transforming the EA into an optimal transport problem, optimal alignment is achieved. Additionally, we propose a confident entity pair generation method based on the Optimal transport algorithm for unsupervised learning. Numerous experiments on four datasets have shown that ELTEA outperforms the previous methods in both supervised and unsupervised situations and is less time-costly.
Journal
IF:
15.5
Papers:
4.1K
Citations:
2.7W
Organization
No organization information available