End-to-End Structure-Aware Convolutional Networks for Knowledge Base Completion.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Knowledge graph embedding is crucial for knowledge base completion.
• Existing models like ConvE excel but lack structure enforcement.
• Graph Convolutional Networks (GCN) leverage graph connectivity for node embeddings.

Purpose of the Study:

• To propose a novel Structure-Aware Convolutional Network (SACN) integrating GCN and ConvE benefits.
• To enhance knowledge graph embedding accuracy by enforcing structure in the embedding space.
• To improve link prediction performance in knowledge graphs.

Main Methods:

• Developed a Structure-Aware Convolutional Network (SACN) with a weighted graph convolutional network (WGCN) encoder and a Conv-TransE decoder.
• WGCN incorporates node structure, attributes (as additional nodes), and edge relation types with learnable weights for adaptive neighbor aggregation.
• Conv-TransE decoder ensures translational properties while maintaining ConvE's link prediction performance.

Main Results:

• The proposed SACN demonstrated effectiveness on FB15k-237 and WN18RR datasets.
• Achieved approximately 10% relative improvement over the state-of-the-art ConvE.
• Outperformed ConvE in HITS@1, HITS@3, and HITS@10 metrics.

Conclusions:

• SACN successfully integrates graph structure awareness with convolutional neural networks for knowledge graph embedding.
• The model offers superior performance in knowledge base completion tasks compared to existing methods.
• This approach provides a scalable and accurate solution for learning embeddings in large knowledge graphs.