Learning Knowledge Graph Embedding With Heterogeneous Relation Attention Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Knowledge graph (KG) embedding seeks to represent KGs in a lower-dimensional space while preserving their structure.
• Graph neural networks (GNNs) are powerful tools for graph representation learning.
• Heterogeneous KGs, with diverse entities and relations, pose challenges for standard GNNs in aggregating multi-semantic information.

Purpose of the Study:

• To propose a novel heterogeneous GNN framework that effectively handles complex graph structures and aggregates multi-semantic information.
• To develop a method that can capture various types of semantic information and selectively aggregate informative features for KG embedding.

Main Methods:

• A heterogeneous GNN framework employing an attention mechanism is proposed.
• Neighbor features are aggregated under each relation-path.
• Relation features are utilized to learn the importance of different relation-paths.
• Weighted aggregation of relation-path-based features generates the final embedding representation.

Main Results:

• The proposed method successfully aggregates entity features from different semantic aspects.
• It selectively aggregates informative features by assigning appropriate weights.
• Experiments on three real-world KGs show superior performance compared to state-of-the-art methods.

Conclusions:

• The novel heterogeneous GNN framework effectively addresses the challenge of KG embedding in heterogeneous graphs.
• The attention-based approach enables selective aggregation of multi-semantic information, leading to improved embedding representations.
• This work offers a promising direction for advancing KG embedding techniques.