Pair-wise or high-order? A self-adaptive graph framework for knowledge graph embedding.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Knowledge graphs (KGs) are crucial for AI but suffer from incompleteness.
• Knowledge graph embedding (KGE) addresses this by representing entities and relations as vectors.
• Graph convolutional networks (GCNs) are key KGE models but face challenges like over-smoothing and limited feature integration.

Purpose of the Study:

• To address limitations in GCNs for knowledge graph representation learning.
• To develop a model that effectively integrates pair-wise and high-order features.
• To improve the accuracy and performance of knowledge graph embeddings.

Main Methods:

• Proposed PHGCN (Pair-wise and High-Order Graph Convolutional Network), an adaptive GCN model.
• Utilized a layer-aware GCN to mitigate over-smoothing in pair-wise relationship aggregation.
• Employed simplicial complex neural networks to extract high-order topological features.
• Introduced a self-adaptive aggregation mechanism for integrating diverse features.

Main Results:

• PHGCN achieved state-of-the-art results on four benchmark datasets.
• Demonstrated significant performance improvements due to high-order feature extraction.
• Achieved a 1.5% improvement on FB15k-237 and 6.1% on WN18RR.

Conclusions:

• PHGCN effectively overcomes over-smoothing and integrates pair-wise and high-order features.
• The model offers superior performance in knowledge graph embedding tasks.
• Highlights the importance of high-order topological features for enhanced KG representation.