Enhancing Recommender Systems with Generative LLM-Based Reinforcement Learning Agents for Data-Efficient Personalization

Abstract

Shirmohammad Tavangari and Asef Yelghi

Recommender systems are important for online platforms, but they suffer from data shortages, cold start issues, and lack of real-time personalization. This paper presents a hybrid framework for building adaptive and data-optimized recommender systems that utilizes large language models and reinforcement learning. The large language model generates high-quality synthetic data, and the reinforcement learning agent improves suggestions with user feedback; a self-tuning mechanism is also used to select data and accelerate learning. Experiments on real-world datasets (Movie Lens and Amazon Reviews) under low-data conditions show that our model significantly outperforms classical and neural baselines, including Matrix Factorization, Deep MF, and DDPG-based RS, in terms of recommendation accuracy, convergence speed, and personalization quality.

PDF