.png){kind=icon}
.png){kind=icon}
Return
A mechanistic study on the impact of entity degree distribution in open-world link prediction
.png){kind=icon}
PRE
AI
D
X
Y
D
DOI：10.1016/j.ipm.2025.104565.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Current research in open-world link prediction typically attributes performance limitations to insufficient entity representation and a lack of entity-relation interaction. Consequently, most proposed improvements focus on addressing these specific issues. However, this research paradigm has yielded limited gains on benchmark datasets with stricter standards, such as FB15K-237-OWE, where inverse relations are explicitly removed. This outcome indicates that existing methods have not yet overcome the constraints imposed by the dataset’s structural characteristics. The underlying cause appears to be a lack of insight into how these structural features fundamentally influence model performance. Therefore, this study focuses on the FB15K-237-OWE knowledge graph dataset to explore the mechanisms by which its structural characteristics affect open-world link prediction performance. First, by designing a subgraph sampling method and conducting Sobol sensitivity analysis, we demonstrate the significant impact of entity degree distribution on model performance. Second, correlation analysis reveals a positive correlation between entity degree and prediction performance. Furthermore, this study investigates how entity link degree influences embedding space distribution and weight updates during neural network training, uncovering its deep impact on performance. Finally, to validate the utility of this mechanistic insight, we designed a targeted experiment to mitigate the imbalance caused by the identified gradient dominance effect. This approach yielded improvements of 2 % to 3 % across all metrics. This study lays the foundation for targeted improvements in open-world link prediction models.
Journal
I
IF:
6.9
Papers:
5.2K
Citations:
1.4W
Organization
No organization information available
Cited Papers