Feature-decorrelation adaptive contrastive learning for knowledge-aware recommendation.

Area of Science:

• Artificial Intelligence
• Data Science
• Recommender Systems

Background:

• Knowledge graphs (KGs) offer rich semantic information for recommendation systems.
• Graph neural networks (GNNs) capture multi-hop relationships but face challenges with complex entity features and irrelevant knowledge propagation.
• Existing GNN methods can suffer from feature loss, distortion, and topic deviation in recommendations.

Purpose of the Study:

• To address limitations in GNN-based knowledge-aware recommendation systems.
• To propose a method that mitigates feature loss and irrelevant knowledge impact.
• To enhance the accuracy and relevance of recommendations by improving knowledge representation.

Main Methods:

• Developed a feature-decorrelation adaptive contrastive learning method.
• Introduced a constraint method to learn representations by investigating inter-feature correlations.
• Proposed an adaptive knowledge refinement technique to extract high-order semantics and generate augmented views.
• Implemented a contrastive learning approach to focus representations on the recommended topic.

Main Results:

• The proposed method effectively models complex entity features and refines knowledge.
• Feature decorrelation significantly improves GNN-based knowledge-aware recommender systems.
• Experiments on Movielens and Yelp datasets validate the method's effectiveness.
• The approach successfully reduces the negative impact of irrelevant knowledge.

Conclusions:

• The feature-decorrelation adaptive contrastive learning method offers a robust solution for knowledge-aware recommendations.
• This technique enhances knowledge representation and recommendation accuracy.
• The findings suggest a promising direction for future research in GNN-based recommender systems.