Regularizing Knowledge Transfer in Recommendation With Tag-Inferred Correlation.

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Traditional recommender systems face challenges with data sparsity.
• Cross-domain recommender systems leverage user knowledge from auxiliary domains to improve recommendations in a target domain.
• Effective domain correlation is crucial for successful knowledge transfer in cross-domain recommendations.

Purpose of the Study:

• To propose a complete tag-induced cross-domain recommendation (CTagCDR) model.
• To fully exploit knowledge from both shared and domain-specific social tags for improved recommendations.
• To address limitations of existing models that only use a subset of shared tags.

Main Methods:

• Developed the CTagCDR model to infer interdomain and intradomain correlations from tagging history.
• Applied learned structural constraints to regularize joint matrix factorization.
• Utilized social tags to explicitly link different domains, even without user or item overlap.

Main Results:

• CTagCDR demonstrated strong performance on rating prediction and item recommendation tasks.
• The model effectively improved recommendation performance compared to state-of-the-art approaches.
• Evaluated on three public datasets against five single and cross-domain recommendation methods.

Conclusions:

• The proposed CTagCDR model successfully addresses data sparsity by fully utilizing tag information.
• CTagCDR offers an effective paradigm for cross-domain recommendation by enhancing knowledge transfer.
• The model's ability to leverage comprehensive tag data leads to superior recommendation outcomes.