Modeling the Influence of Content Recommender Systems on Topic Diffusion in Online Social Network Simulations

This thesis employs a controlled simulation environment to investigate how recommender systems affect the way topics spread in online social networks. We developed a rich and dynamic social ecosystem by expanding upon the YSocial platform, a digital twin populated by agents driven by Large Language Models (LLMs). In this ecosystem, agents engage, post, and follow one another according to the content they come across. From interest-based models to collaborative filters based on user similarity, our work concentrated on developing and putting into practice various recommender strategies. We developed new customized metrics, such as personalization balance, engagement momentum, sentiment diffusion over time in order to methodically assess content recommenders effects. We examined the ways in which recommendation strategies influence topic diffusion through comprehensive experiments. The results highlight the magnitude of different algorithmic design choices in shaping the structure and diversity of online network conversations.