MRGAT: Multi-Relational Graph Attention Network for knowledge graph completion.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Embedding-based models are effective for knowledge graph completion.
• Graph neural networks (GNNs) leverage structural information but assume equal neighbor importance.
• Heterogeneous knowledge graphs with multiple relations pose challenges for GNN message passing.

Purpose of the Study:

• To address the limitations of current GNN-based knowledge graph completion methods.
• To propose a novel model that accounts for varying neighbor importance and complex relational interactions.
• To enhance the performance of knowledge graph completion.

Main Methods:

• Designed a multi-relational graph attention network (MRGAT).
• Incorporated a self-attention layer to calculate the importance of neighboring nodes.
• Adapted the network to handle heterogeneous multi-relational connections.

Main Results:

• MRGAT demonstrated superior performance on benchmark knowledge graphs.
• Achieved state-of-the-art results across various evaluation metrics, including MRR and Hits@ scores.
• Validated the effectiveness of assigning different weights to neighboring nodes.

Conclusions:

• The assumption of equal neighbor importance in GNNs is unreasonable for knowledge graph completion.
• MRGAT effectively optimizes network structure by incorporating self-attention and differential node weighting.
• The proposed model significantly improves knowledge graph completion performance.