.png){kind=icon}
.png){kind=icon}
Return
Unified link prediction modeling for enhanced knowledge graph completion task
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127356.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Link prediction in knowledge graphs (KGs) aims to identify missing links between entities. Existing studies primarily focus on specific scenarios, such as transductive (seen-to-seen) or inductive (seen-to-unseen, unseen-to-seen, or unseen-to-unseen) settings, individually. However, real-world challenges arise when both unseen entities and unseen relations appear simultaneously during testing, creating a more complex and realistic scenario. To address this gap, we propose a unified method with three key components designed to enhance adaptability and accuracy across diverse link prediction settings: (1) entity-independent modeling with triple-view graph, which employs graph neural networks (GNNs) to learn relation patterns independently of entities, enabling more effective inductive knowledge graph completion; (2) contrastive learning-based relation-context modeling, which mitigates the lack of connectivity information in KGs by allowing GNNs to propagate and learn meaningful representations for unseen entities; and (3) unseen relations modeling with augmented schema, which leverages the KG's ontological schema to flexibly model unseen relations and predict missing links without requiring extensive retraining. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of our framework in terms of accuracy and adaptability across various scenarios, outperforming state-of-the-art baselines and underscoring its potential for addressing real-world knowledge graph completion challenges.
Keywords:
Knowledge graph completion
Graph neural network
Link prediction
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
View more organizations