.png){kind=icon}
.png){kind=icon}
Return
Hyperparameter-Free Bi-level Knowledge Graph Optimization for Link Prediction
.png){kind=icon}
PRE
AI
DOI：10.1007/978-3-032-04549-2_2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In this paper, we explore the challenges and advancements in Bi-level Knowledge Graphs, which extend traditional triplet-based representations to capture relationships between facts. While Bi-level KGs offer greater expressiveness, they encounter issues related to sparsity and the introduction of hyperparameters in data augmentation modules, designed to enrich the graph structure. Specifically, the hyperparameters tau and lambda, which control the scale of augmented data and its loss weight, have been set empirically without sufficient exploration of their optimal values across different dataset sizes. To address this, we propose a method that eliminates these redundant hyperparameters by introducing functions F(.) and G(.) to determine the optimal scale of augmented data and loss weight. Our approach uses KL divergence to ensure alignment between the augmented and original datasets, and identifies a quadratic relationship between model performance and the loss weight. The effectiveness of our method is validated on three datasets, demonstrating significant improvements in triplet prediction and conditional link prediction tasks compared to existing models. This strategy not only improves model performance but also reduces reliance on empirical tuning, leading to more efficient and scalable knowledge graph reasoning models.
Keywords:
Knowledge graph
Parameter elimination
Link prediction
Journal
A
IF:
0
Papers:
24
Citations:
0
View more organizations