Simulating Filter Bubble on Short-video Recommender System with Large Language Model Agents

TL;DR

This work investigates how filter bubbles emerge in short-video recommender systems by deploying a large language model (LLM)–based agent framework integrated with matrix factorization and factorization machines–based recommenders and real-world video data. The closed-loop simulation enables realistic user–feedback dynamics, revealing that demographic features and category attraction influence content homogenization and that interventions such as cold-start matching and progressive feedback weighting can mitigate bubble effects while preserving user satisfaction. Empirical results show that personality-driven motivations tend to produce stronger bubbles, whereas uses-and-gratifications motivations yield more diverse exposure; demographic signals like age and phone price also modulate diversity. The framework offers a rapid prototyping platform for designing diverse, inclusive recommender strategies and highlights practical considerations for reducing algorithmic bias in digital spaces.

Abstract

An increasing reliance on recommender systems has led to concerns about the creation of filter bubbles on social media, especially on short video platforms like TikTok. However, their formation is still not entirely understood due to the complex dynamics between recommendation algorithms and user feedback. In this paper, we aim to shed light on these dynamics using a large language model-based simulation framework. Our work employs real-world short-video data containing rich video content information and detailed user-agents to realistically simulate the recommendation-feedback cycle. Through large-scale simulations, we demonstrate that LLMs can replicate real-world user-recommender interactions, uncovering key mechanisms driving filter bubble formation. We identify critical factors, such as demographic features and category attraction that exacerbate content homogenization. To mitigate this, we design and test interventions including various cold-start and feedback weighting strategies, showing measurable reductions in filter bubble effects. Our framework enables rapid prototyping of recommendation strategies, offering actionable solutions to enhance content diversity in real-world systems. Furthermore, we analyze how LLM-inherent biases may propagate through recommendations, proposing safeguards to promote equity for vulnerable groups, such as women and low-income populations. By examining the interplay between recommendation and LLM agents, this work advances a deeper understanding of algorithmic bias and provides practical tools to promote inclusive digital spaces.

Figures (9)

• Figure 1: Item data statistics.
• Figure 2: The overview of our simulation framework, which is composed of two major components, the LLM module, which acts as a user-agent, and the recommender module, which generates video suggestions for the user to react to.
• Figure 3: Coverage and satisfaction over time for users motivated by uses and gratifications.
• Figure 4: Filter bubble changes over time for three category levels for users motivated by uses and gratifications.
• Figure 5: Coverage and satisfaction over time for users motivated by personality.