Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AgentCF: Collaborative Learning with Autonomous Language Agents for Recommender Systems

Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian McAuley, Wayne Xin Zhao, Leyu Lin, Ji-Rong Wen

TL;DR

AgentCF tackles the gap in simulating recommender-system user behavior by treating both users and items as autonomous language-model agents equipped with memory. The framework uses autonomous interactions plus collaborative reflection to align agent decisions with real interaction records, enabling implicit two-sided collaborative filtering and preference propagation. Empirical results on Amazon CDs and Office subsets show that AgentCF variants achieve competitive or superior ranking performance with far less data than traditional models, while producing richer, more personalized interaction behaviors including user-user and item-item dynamics. The work demonstrates a path toward more realistic user-behavior simulations and potential for AI-driven agent ecosystems in recommender systems.

Abstract

Recently, there has been an emergence of employing LLM-powered agents as believable human proxies, based on their remarkable decision-making capability. However, existing studies mainly focus on simulating human dialogue. Human non-verbal behaviors, such as item clicking in recommender systems, although implicitly exhibiting user preferences and could enhance the modeling of users, have not been deeply explored. The main reasons lie in the gap between language modeling and behavior modeling, as well as the incomprehension of LLMs about user-item relations. To address this issue, we propose AgentCF for simulating user-item interactions in recommender systems through agent-based collaborative filtering. We creatively consider not only users but also items as agents, and develop a collaborative learning approach that optimizes both kinds of agents together. Specifically, at each time step, we first prompt the user and item agents to interact autonomously. Then, based on the disparities between the agents' decisions and real-world interaction records, user and item agents are prompted to reflect on and adjust the misleading simulations collaboratively, thereby modeling their two-sided relations. The optimized agents can also propagate their preferences to other agents in subsequent interactions, implicitly capturing the collaborative filtering idea. Overall, the optimized agents exhibit diverse interaction behaviors within our framework, including user-item, user-user, item-item, and collective interactions. The results show that these agents can demonstrate personalized behaviors akin to those of real-world individuals, sparking the development of next-generation user behavior simulation.

AgentCF: Collaborative Learning with Autonomous Language Agents for Recommender Systems

TL;DR

AgentCF tackles the gap in simulating recommender-system user behavior by treating both users and items as autonomous language-model agents equipped with memory. The framework uses autonomous interactions plus collaborative reflection to align agent decisions with real interaction records, enabling implicit two-sided collaborative filtering and preference propagation. Empirical results on Amazon CDs and Office subsets show that AgentCF variants achieve competitive or superior ranking performance with far less data than traditional models, while producing richer, more personalized interaction behaviors including user-user and item-item dynamics. The work demonstrates a path toward more realistic user-behavior simulations and potential for AI-driven agent ecosystems in recommender systems.

Abstract

Recently, there has been an emergence of employing LLM-powered agents as believable human proxies, based on their remarkable decision-making capability. However, existing studies mainly focus on simulating human dialogue. Human non-verbal behaviors, such as item clicking in recommender systems, although implicitly exhibiting user preferences and could enhance the modeling of users, have not been deeply explored. The main reasons lie in the gap between language modeling and behavior modeling, as well as the incomprehension of LLMs about user-item relations. To address this issue, we propose AgentCF for simulating user-item interactions in recommender systems through agent-based collaborative filtering. We creatively consider not only users but also items as agents, and develop a collaborative learning approach that optimizes both kinds of agents together. Specifically, at each time step, we first prompt the user and item agents to interact autonomously. Then, based on the disparities between the agents' decisions and real-world interaction records, user and item agents are prompted to reflect on and adjust the misleading simulations collaboratively, thereby modeling their two-sided relations. The optimized agents can also propagate their preferences to other agents in subsequent interactions, implicitly capturing the collaborative filtering idea. Overall, the optimized agents exhibit diverse interaction behaviors within our framework, including user-item, user-user, item-item, and collective interactions. The results show that these agents can demonstrate personalized behaviors akin to those of real-world individuals, sparking the development of next-generation user behavior simulation.
Paper Structure (41 sections, 4 equations, 8 figures, 3 tables)

This paper contains 41 sections, 4 equations, 8 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Preliminaries
  4. Collaborative Agent Optimization
  5. Memory Design
  6. Autonomous Interactions for Contrastive Item Selection
  7. Collaborative Reflection and Memory Update
  8. Connection with Classical Recommendation Models
  9. Agent Interaction Inference
  10. Experiments
  11. Experimental Setup
  12. Datasets
  13. Evaluation Metrics
  14. Baseline Models
  15. Overall Performance
  16. ...and 26 more sections

Figures (8)

  • Figure 1: The overall framework of AgentCF and a case about the optimization process of agents: (1) The user and item agents are first prompted to autonomously interact. (2) These agents adjust the misconceptions in their memory, by reflecting on the disparities between their decisions and real-world interactions. In this process, the simulated preferences of user and item agents aggregate (as indicated by the highlighted content) and can propagate to other agents in subsequent interactions.
  • Figure 2: Analysis of whether our approach can simulate personalized agents to mitigate position bias and popularity bias.
  • Figure 3: Performance comparison w.r.t. the progress of optimization. The Y-axis denotes the proportion of agents making accurate choices. The X-axis denotes the step of optimization. "Test" indicates the results in the test dataset.
  • Figure 4: Proportion of agents buying items after viewing reviews. "Similar" means that the reviews are written by users with similar preferences to test users. "Neg" indicates that the review is negative.
  • Figure 5: Performance comparisons w.r.t. cold-start item agents' interactions with different popular item agents. "Dist" means that the interacted item has distinct identity information compared to the cold-start item. "Sim" means similar.
  • ...and 3 more figures