Community knowledge graph abstraction for enhanced link prediction: A study on PubMed knowledge graph.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Natural Language Processing

Background:

• Biomedical knowledge is rapidly expanding, overwhelming manual updates for existing Knowledge Graphs (KGs).
• Existing Knowledge Graph Embedding (KGE) methods struggle to capture attribute-specific entity relationships in KGs.

Purpose of the Study:

• To propose an extension approach for KGE models to improve link prediction in biomedical KGs.
• To address the limitations of current KGE methods in capturing complex entity relationships.

Main Methods:

• Developed an entity distance-based method to abstract Community Knowledge Graphs (CKGs) from the PubMed Knowledge Graph (PKG).
• Extended existing KGE models by integrating PKG information into the abstracted CKGs for link prediction.
• Evaluated performance using TransE, TransH, DistMult, ComplEx, SimplE, and RotatE models with metrics like MR, MRR, and Hits@k.

Main Results:

• The proposed extension improved link prediction accuracy across all six evaluated KGE models.
• Top-10 accuracy increased significantly, for example, from 0.76 to 0.85 for RotatE.
• Results demonstrate the broad applicability and effectiveness of the extension approach.

Conclusions:

• The study presents a novel method for abstracting CKGs and enhancing KGE performance.
• The extension approach shows significant improvements in link prediction for biomedical KGs.
• Future work includes applying link prediction to newly introduced entities in the PKG.