Multitype view of knowledge contrastive learning for recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Graph Neural Networks (GNNs) are crucial for recommender systems, leveraging diverse knowledge and interaction data.
• Contrastive learning enhances GNNs by providing supervisory signals, but current methods have limitations in knowledge graph integration.
• Effective knowledge graph utilization is key to overcoming challenges like data sparsity and noise in recommendations.

Purpose of the Study:

• To propose a novel multitype view of knowledge contrastive learning for recommendations (MVKC) model.
• To enhance the knowledge perception capabilities of GNN-based recommender systems.
• To address limitations in existing contrastive learning settings for knowledge graph-enhanced recommendations.

Main Methods:

• The MVKC model employs two modules to generate hierarchical and augmented views of the data.
• Hierarchical views capture global and local knowledge at multiple levels.
• Augmented views are created by fusing augmented knowledge and interaction graphs, enabling self-supervised feature mining.

Main Results:

• The MVKC model effectively alleviates user-item interaction graph sparsity and suppresses knowledge graph noise.
• It successfully filters long-tail entities, improving recommendation quality.
• Experimental results on three public datasets show MVKC outperforms state-of-the-art methods.

Conclusions:

• The proposed MVKC model offers a robust and effective approach to knowledge graph-enhanced recommendations.
• Its multitype view contrastive learning strategy significantly improves GNN performance in recommender systems.
• MVKC demonstrates strong anti-interference ability and robustness, crucial for practical applications.