Paper logo

Paper

Free

LLM-empowered generative agents for simulating user behavior in recommender systems

FreeFree tier
Type
Open Source

About Paper

Agent4Rec is a user simulator for recommender systems that leverages large language model (LLM) empowered generative agents. It is designed to bridge the gap between offline metrics and online performance in recommendation research. Each agent is equipped with a user profile module initialized from real-world datasets (e.g., MovieLens, Steam, Amazon-Book), a memory module that logs factual and emotional memories with an emotion-driven reflection mechanism, and an action module supporting both taste-driven and emotion-driven behaviors. Agents interact with personalized recommender models in a page-by-page manner using a pre-implemented collaborative filtering algorithm. The framework enables multi-faceted evaluation of agent alignment with real user preferences, emulation of filter bubble effects, and discovery of causal relationships in recommendation tasks. The code is openly available.

Key Features

LLM-empowered generative agents with user profile, memory, and action modules
User profile initialization from real-world datasets (MovieLens, Steam, Amazon-Book)
Emotion-driven reflection mechanism in memory module
Taste-driven and emotion-driven action support
Page-by-page interaction with personalized recommender models
Collaborative filtering-based recommendation algorithm
Multi-faceted evaluation of agent-user alignment
Capability to emulate filter bubble effects and discover causal relationships

Pros & Cons

Pros
  • Open source with publicly available code
  • Leverages state-of-the-art LLMs for human-level intelligence
  • Initialized from real-world datasets for realistic user profiles
  • Supports both taste-driven and emotion-driven actions
  • Enables multi-faceted evaluation and insightful experiments like filter bubble emulation
  • Published as a SIGIR 2024 perspective paper with rigorous evaluation
Cons
  • Current implementation limited to collaborative filtering-based recommendation algorithm
  • May not fully capture the full complexity of real human behavior
  • Dependency on LLM quality and prompt engineering for agent behavior

Best For

Simulating user behavior in recommender systems for offline evaluationStudying the impact of filter bubbles on user experienceDiscovering causal relationships in recommendation tasksTesting and comparing recommender algorithms in a controlled environmentResearch on human-like generative agents in information systems