High-Order Neighbors Aware Representation Learning for Knowledge Graph Completion.

Area of Science:

• Artificial Intelligence
• Data Science
• Computer Science

Background:

• Knowledge graph completion (KGC) is crucial for knowledge acquisition, aiming to infer missing facts in knowledge graphs (KGs).
• Existing graph convolutional network (GCN)-based knowledge graph embedding (KGE) methods struggle with predicting distant or unreachable entities due to ignored high-order information.

Purpose of the Study:

• To enhance KGC performance by effectively learning from high-order neighbors in knowledge graphs.
• To address the limitations of current KGE methods in capturing distant or unreachable entity relationships.

Main Methods:

• Introduced 'pedalnodes' to augment KGs, facilitating message passing and injecting high-order neighbor information into entity representations.
• Proposed strength-guided graph neural networks for aggregating neighboring entity representations.
• Developed a dynamic integration method to combine aggregated representations with self-representations, mitigating irrelevant information transfer.

Main Results:

• The proposed method significantly improves the prediction of missing triplets in knowledge graphs.
• Experimental results on three benchmark datasets demonstrate the superiority of the approach over strong baseline models.
• The integration of high-order information through pedalnodes and strength-guided aggregation enhances entity representation.

Conclusions:

• The novel approach effectively leverages high-order information for improved knowledge graph completion.
• The method offers a promising solution for overcoming limitations in existing KGE techniques.
• This work contributes to advancing the field of knowledge graph representation learning.