How does Search Affect Personalized Recommendations and User Behavior? Evidence from LLM-based Synthetic Data

Search and personalized recommendations represent two prominent Information Retrieval (IR) technologies often used interchangeably by users across online platforms. Recent scholarly and industrial efforts have shifted towards unraveling the intricate interplay between these technologies within user studies and algorithmic frameworks, aiming to foster a deeper understanding of user behaviors and enhance the precision of user modeling. Nevertheless, due to concerns regarding data privacy and the substantial risks associated with data acquisition, the precise mechanisms by which how search influences personalized recommendations and user behavior remain largely unclear. This investigation adopts a Large Language Model (LLM)-based generative agent simulation methodology to produce synthetic user behavioral data, ensuring its refinement and reliability. Through this approach, we examine issues surrounding search utility and delineate the impact of search functionalities on user behavior in the scenario of online retail, while considering both the longitudinal and cross-sectional heterogeneity of these influences. The insights gleaned from this inquiry provide invaluable perspectives for improving user understanding, refining user modeling methodologies, and designing more efficacious algorithms.