Retrieval Augmented Multi-agent Recommender

Online review platforms are vital for shaping consumer choices and experiences in today's digital world. The AgentSociety Challenge focuses on advancing personalized recommendations with large language model (LLM) agents, aiming to enhance web interactions and user satisfaction. This paper presents the first-place solution for the AgentSociety Challenge Recommendation Track, by the team of the University of Electronic Science and Technology of China. We propose a recommendation framework based on large language models, integrating retrieval-augmented feature extraction and a data-driven multi-agent strategy. The core advantages of our approach lie in two key aspects. First, to optimize the representation of both products and users, we design a retrieval-augmented feature extraction mechanism that combines structured product attributes with unstructured user reviews, capturing both the objective characteristics of products and users' subjective preferences. Second, we introduce a data-driven multi-agent system that adaptively adjusts processing strategies based on the heterogeneous characteristics of different datasets, thereby effectively uncovering valuable latent information. Our method is built upon the Qwen-72B-Instruct model, with further optimizations in recommendation accuracy and system stability. Experimental results on three widely used datasets-Amazon, Goodreads, and Yelp-demonstrate significant improvements in recommendation performance. Ultimately, our solution achieved an overall hit rate score of 0.4744, representing a 101% improvement over the baseline, showcasing its outstanding effectiveness in personalized recommendation tasks.