Meta learning based few shot knowledge graph completion with domain selected aggregation.

Area of Science:

• Artificial Intelligence
• Data Science
• Knowledge Representation

Background:

• Knowledge graphs are crucial for AI reasoning tasks.
• Data sparsity poses a significant challenge for few-shot knowledge graph completion.
• Current methods struggle with noisy neighbors and lack sensitivity to semantic task characteristics.

Purpose of the Study:

• To develop a robust few-shot knowledge graph completion method addressing data sparsity and noise.
• To enhance the expressiveness and task-awareness of relation representations.
• To improve the adaptability of models to new knowledge graph completion tasks.

Main Methods:

• Proposed a domain-selected neighborhood aggregation mechanism to filter irrelevant entities.
• Introduced a relation meta-learner integrating contextual attention and multi-layer perceptron.
• Employed a meta-optimization strategy for rapid adaptation via an embedding learner.

Main Results:

• The proposed method significantly outperforms state-of-the-art baselines on NELL-One and Wiki-One datasets.
• Achieved notable performance improvements on the Hits@10 metric in 5-shot tasks.
• Demonstrated effective noise suppression and improved relation inference under sparse conditions.

Conclusions:

• The meta-learning based approach with domain-selected aggregation effectively tackles few-shot knowledge graph completion challenges.
• The method generates more expressive, task-aware relation representations, enhancing inference.
• The approach shows strong generalization capabilities and adaptability to new tasks.