Year: 2025
Type: conference-paper
Abstract: With powerful semantic understanding, planning, and decision-making capabilities, LLM-based agents have been applied to solve recommendation tasks. However, challenges remain in effectively evolving user preferences within agent modules and in leveraging external knowledge to improve recommendation accuracy. In this paper, we propose an innovative agent-based framework for personalized recommendation systems that integrates adaptive knowledge fusion to address these challenges. Our framework consists of two primary components: an intelligent agent and the fusion of external knowledge. The intelligent agent includes a memory module that simulates human-like retention of user-item interactions and a reasoning module that applies Chain-of-Thought (CoT) techniques to mimic human thought processes. Furthermore, we introduce two main strategies for knowledge fusion: (1) a preranking approach using external knowledge to reorder candidates and reduce bias, and (2) an ensemble method that combines multiple ranking signals for more accurate recommendations. Evaluations on three public datasets and the AgentSociety Challenge demonstrate the framework's effectiveness in enhancing recommendation quality and adaptability. The code is open-sourced.
Authors: Yuanqing Yu, Zhefan Wang, Chumeng Jiang, Xinyi Li, Jiayin Wang, Min Zhang
Institutions: Quan Cheng Laboratory, Tsinghua University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Award: G3379573397
OpenAlex ID: W4410636803