MuLAN: Multi-level attention-enhanced matching network for few-shot knowledge graph completion.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Few-shot knowledge graph completion (KGC) is crucial for augmenting knowledge graphs with limited relation data.
• Current methods often rely on entity embeddings from one-hop neighbors, neglecting crucial inter-neighbor dynamics and feature importance.

Purpose of the Study:

• To address limitations in existing few-shot KGC methods, particularly regarding inter-neighbor interactions and feature dimension significance.
• To propose a novel network, MuLAN, for enhanced few-shot knowledge graph completion.

Main Methods:

• Developed a Multi-Level Attention-enhanced matching Network (MuLAN) incorporating a multi-head self-attention neighbor encoder.
• Implemented entity-level, instance-level, and feature-level attention mechanisms for comprehensive matching.
• Introduced a consistency constraint to enhance the stability of support instance embeddings.

Main Results:

• MuLAN demonstrated significant advantages over 11 state-of-the-art competitors on NELL-One and Wiki-One datasets.
• Achieved an average improvement of 14.5% in MRR and 13.3% in Hits@K compared to the best baseline.
• Effectively handles inter-neighbor interaction, local entity matching, and varying feature dimension significance.

Conclusions:

• MuLAN offers a superior approach to few-shot knowledge graph completion by effectively modeling complex relationships and feature importance.
• The proposed attention mechanisms and consistency constraint lead to substantial performance gains in KGC tasks.
• This work advances the field of few-shot learning in knowledge graphs, paving the way for more robust and accurate knowledge graph augmentation.