An inductive knowledge graph embedding via combination of subgraph and type information

Area of Science:

• Artificial Intelligence
• Data Science
• Machine Learning

Background:

• Conventional knowledge graph representation methods project entities and relations into vector spaces, enhancing link prediction and downstream tasks.
• These methods struggle with knowledge graph evolution, failing to process previously unseen entities in target knowledge graphs.
• Existing inductive subgraph-based models for link prediction overlook crucial semantic information.

Purpose of the Study:

• To develop an inductive representation learning model capable of handling evolving knowledge graphs with unseen entities.
• To integrate both topological structure and semantic information for more robust knowledge graph representations.
• To improve the performance of link prediction in dynamic knowledge graph environments.

Main Methods:

• Propose TGraiL, an inductive representation learning model for knowledge graphs.
• Encode node topological structure using subgraph distances.
• Encode entity type information via projection matrices.
• Fuse topological and semantic information for training entity vector representations.

Main Results:

• TGraiL demonstrates significantly improved performance compared to existing baseline models.
• The model effectively integrates topological structure and semantic information for enhanced representations.
• Experimental results validate the effectiveness and superiority of the proposed TGraiL method.

Conclusions:

• TGraiL offers an effective solution for inductive representation learning in evolving knowledge graphs.
• The integration of topological and semantic information is crucial for handling unseen entities.
• The proposed method advances the state-of-the-art in knowledge graph representation learning and link prediction.