Dynamic and Static Representation Learning Network for Recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current review-based recommendation systems utilize static strategies, failing to account for the evolving nature of user preferences and item popularity.
• This limitation hinders the accuracy of latent representations for users and items in recommendation models.

Purpose of the Study:

• To propose a Dynamic and Static Representation Learning Network (DSRLN) for improved rating prediction accuracy in recommendation systems.
• To effectively model both the dynamic evolution of user interests and the static intrinsic preferences of users.

Main Methods:

• Developed DSRLN incorporating a dynamic representation extractor to analyze user interaction sequences and capture evolving interests.
• Integrated a static representation extractor to learn intrinsic user preferences from review semantic coherence and feature strength.
• Employed a personalized adaptive fusion module with a weighted attention mechanism to balance dynamic and static feature influences.

Main Results:

• Extensive experiments on five real-world Amazon datasets demonstrated the superiority of the proposed DSRLN model over existing methods.
• Ablation studies confirmed the effectiveness and contribution of individual components within the DSRLN architecture.

Conclusions:

• The DSRLN model significantly enhances rating prediction accuracy by effectively integrating dynamic and static user and item representations.
• The proposed approach offers a more nuanced understanding of user behavior and item characteristics for improved recommendation performance.