• human-AI loop;information operations;polarization

Modern human interactions predominantly occur online due to the increasing digitalization of communication tools. However, these online social dynamics are challenging to measure, primarily because of the large volume of online footprints and the co-shaping effects of algorithms that drive the business revenues of online platforms. Such challenges hinder the ability of non-quantitative researchers, policymakers, and platforms to fully understand human societies, thereby posing significant threats to online users and potentially exposing vulnerabilities to malicious actors.

In the first part of the thesis, we employ simulation models to analyze the influence of recommender systems on individual opinions and preferences, exploring both user-user and user-product interactions driven by network and collaborative-filtering recommender algorithms. In the second part, we develop graph machine learning techniques to provide stakeholders with tools for detecting threats to online users. Specifically, we propose a method to identify echo chambers and two methodologies to uncover online information operations.

This thesis enhance our understanding of online dynamics, providing us with critical insights and tools that are pivotal in designing and implementing effective interventions and policies.