Implicit Stochastic Gradient Descent Method for Cross-Domain Recommendation System.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Previous recommendation systems using matrix factorization collaborative filtering (MFCF) were limited to single domains, struggling with data sparsity and the cold-start problem.
• Leveraging knowledge across domains (domain coherence) can enhance recommendation quality by transferring insights from source to target domains.

Purpose of the Study:

• To develop a cross-domain recommendation system (CDRS) that addresses the limitations of single-domain MFCF.
• To discover and utilize latent features across multiple domains to improve recommendation performance and mitigate the cold-start issue.

Main Methods:

• Applied matrix factorization collaborative filtering (MFCF) to multiple domains within a unified cross-domain recommendation system (CDRS).
• Utilized the implicit stochastic gradient descent algorithm to optimize the objective function for prediction, consolidating matrices from different domains.
• Designed a conceptual framework for CDRS applicable to various industrial recommender scenarios.

Main Results:

• Experimental results on Amazon Food and MovieLens datasets demonstrated significant improvements: 15.2% in computation time and 19.7% in Mean Squared Error (MSE) compared to other methods.
• Achieved a notably lower convergence value for the loss function, indicating improved model efficiency.
• Analysis revealed a dynamic balance between prediction accuracy and computational complexity.

Conclusions:

• The proposed cross-domain recommendation system effectively leverages domain coherence to enhance recommendation quality and address the cold-start problem.
• The method offers a practical and efficient solution for multi-domain recommendation scenarios, balancing accuracy and computational cost.