Presentation Information
[1I4-GS-4a-06]LLM-based Simulation of News Reading Behavior and Analysis of the Impact of Recommendation Algorithms on Reading Diversity
〇Ren Fujiki1, Yugo Nakamura1, Tsunenori Mine1, Yutaka Arakawa1 (1. Kyushu University)
Keywords:
LLM,RecommenderSystems,Simulation,Information Health
With the advancement of recommendation systems, users news consumption has become increasingly biased toward specific perspectives, raising concerns about filter bubbles and social polarization. While "Information Health," which promotes balanced information exposure, is gaining attention, longitudinal evaluation of recommendation algorithms on real-world platforms remains difficult due to ethical and privacy constraints.
To address this, we developed a news reading simulation environment using Large Language Models (LLMs) and quantitatively analyzed how different recommendation methods affect the temporal evolution of reading diversity and user satisfaction. We had the opportunity to use a Japanese news dataset to generate virtual users with personas derived from browsing history and conducted a six-month simulation.
The experimental results showed that diversity trajectories vary significantly across recommendation methods, and users with strong topic preferences exhibit a pronounced trade-off between diversity and satisfaction. Additionally, personality traits may influence system evaluation.
This study attempts to establish a framework for longitudinally evaluating recommendation systems from the perspective of Information Health, even in the absence of real user logs.
To address this, we developed a news reading simulation environment using Large Language Models (LLMs) and quantitatively analyzed how different recommendation methods affect the temporal evolution of reading diversity and user satisfaction. We had the opportunity to use a Japanese news dataset to generate virtual users with personas derived from browsing history and conducted a six-month simulation.
The experimental results showed that diversity trajectories vary significantly across recommendation methods, and users with strong topic preferences exhibit a pronounced trade-off between diversity and satisfaction. Additionally, personality traits may influence system evaluation.
This study attempts to establish a framework for longitudinally evaluating recommendation systems from the perspective of Information Health, even in the absence of real user logs.