Retrieval-augmented and feedback-optimized large language model for recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Large Language Models (LLMs) offer powerful generative capabilities for recommendation systems (LLM-as-RS).
• Existing LLM-as-RS face challenges including data noise, static models, and input length constraints, leading to costly updates.
• These limitations hinder optimal performance and adaptability in dynamic environments.

Purpose of the Study:

• To propose an innovative generative recommendation framework, KDRAG-Critic-LLM-RS, addressing the limitations of current LLM-as-RS.
• To integrate retrieval augmentation and feedback optimization for enhanced recommendation quality and adaptability.
• To provide a modular and plug-and-play solution for LLM-based recommendation systems.

Main Methods:

• Developed an efficient KDTreeRAG retrieval module using user-based collaborative filtering to identify similar users and provide context.
• Introduced a lightweight R-Critic feedback module inspired by item-based collaborative filtering to refine recommendations using user feedback.
• Established a closed-loop retrieval-generation-feedback mechanism to continuously optimize recommendation quality by integrating user and item perspectives.

Main Results:

• KDRAG-Critic-LLM-RS significantly outperformed traditional baselines in accuracy, personalization, and diversity on movie and book datasets.
• The framework demonstrated effectiveness without requiring any fine-tuning of the Large Language Model.
• Achieved enhanced system responsiveness to dynamic user preferences and improved generalization capabilities.

Conclusions:

• The KDRAG-Critic-LLM-RS framework offers a robust solution for LLM-based recommendation systems, overcoming key limitations.
• Its modular design allows easy integration, avoiding high retraining costs and mitigating data noise.
• The system effectively enhances recommendation quality, adaptability, and generalization by leveraging collaborative signals.