Enhanced knowledge graph recommendation algorithm based on multi-level contrastive learning

Area of Science:

• Artificial Intelligence
• Computer Science
• Data Science

Background:

• Knowledge Graphs (KGs) improve recommendation systems but suffer from data sparsity due to long-tail distributions.
• Existing methods struggle to effectively leverage KG information for personalized recommendations.

Purpose of the Study:

• To propose a knowledge-aware recommendation framework addressing data sparsity in KGs.
• To enhance recommendation accuracy and personalization through multi-level contrastive learning.

Main Methods:

• A Collaborative Knowledge Graph (CKG) is enhanced using random edge dropout for multi-level feature representation (user-user, item-item, user-item).
• Graph Attention Networks (GAT) with a dynamic attention mechanism model the KG.
• Multi-level contrastive learning, incorporating nonlinear transformation and Momentum Contrast (Moco), is used as an auxiliary self-supervised task.

Main Results:

• The framework effectively extracts high-quality feature information from the KG.
• Experimental results on MovieLens and Amazon-books datasets show significant improvements in recommendation performance.
• The proposed method outperforms baseline models across various evaluation metrics.

Conclusions:

• The knowledge-aware recommendation framework successfully mitigates data sparsity issues in KGs.
• Multi-level contrastive learning enhances the effectiveness of KG-based recommendation systems.
• The approach offers a robust solution for personalized recommendations in sparse data environments.