Multipath Cross Graph Convolution for Knowledge Representation Learning.

Area of Science:

• Artificial Intelligence
• Data Science
• Graph Neural Networks

Background:

• Traditional embedding methods struggle with local relationships, hindering end-to-end training.
• Existing models often lack efficient extraction of local entity and relationship features.

Purpose of the Study:

• To propose an end-to-end knowledge graph embedding method that addresses limitations in local relationship training.
• To enhance entity and relation prediction accuracy and reduce computational complexity.

Main Methods:

• Introduced TransC graph convolutional network (TransC-GCN) utilizing local graph convolution and global cross-learning.
• Employed a translation model to map local entities/relationships into a cross vector for GCN input.
• Utilized maximum pooling to focus on strong local features and avoid global redundancy.

Main Results:

• TransC-GCN accurately obtains local entity feature information via lightweight convolutional neural network operations.
• The method achieves significant improvements: 8.1% in entity inference and 4.4% in relation inference accuracy.
• Demonstrated reduced computational complexity and enhanced robustness compared to baseline models.

Conclusions:

• TransC-GCN effectively extracts local features from knowledge graphs, enabling multilayer penetration and relationship derivation.
• The proposed method offers a robust and computationally efficient solution for knowledge graph embedding and prediction.