• computational social science
• mental health communities
• natural language processing
• network science
• online social media
• polluted information environments
• sociopolitical discourse
• supportive information environments

Nowadays, a considerable fraction of social interactions take place on online social platforms that facilitate instantaneous discussions on any kind of information. Several studies underlined that online information environments are polluted, i.e., users' cognitive biases are exacerbated, giving rise to realities that cause alarming phenomena such as extreme partisanship, opinion polarization, and the rise of echo chambers. At the same time, the homophilic mechanisms that drive polluted realities may have different and potentially beneficial effects when observed in other contexts, e.g., supportive groups can form around people who are facing particular challenges, asking for and offering support and assistance.
This thesis investigates this potential duality of online discourse by analyzing both content (opinions, emotional expressions) and network (social interactions) data from social media platforms. Through a multidisciplinary approach combining network science and natural language processing with insights from sociology and psychology, we aim to: i) detect and characterize emergent phenomena in polluted and supportive online spaces, ii) analyze user behaviors and group dynamics within these environments, and iii) quantify their effects across different platforms. Our analysis focuses on two domains that illustrate this potential duality: sociopolitical debates, which may exhibit characteristics of polluted environments, and mental health discussions, which could foster supportive interactions. Through several case studies on Reddit and X/Twitter, we explore how similar homophilic mechanisms can lead to different outcomes at both individual and group levels, from echo chambers in political discourse to mutual aid networks in mental health communities.
Ultimately, we examine how recently introduced limitations on social media data access impact our understanding of these phenomena and propose an LLM-powered digital-twin that could help supplement -- though not replace -- traditional research methods under data constraints.
Our key empirical finding shows that online social platforms can simultaneously foster both potentially harmful and beneficial phenomena, with community norms and interaction patterns, rather than platform architecture alone, playing a crucial role in determining outcomes. From a methodological perspective, we provide researchers with practical, open-source resources for studying online social phenomena while ensuring reproducibility and privacy protection. These findings and tools have significant implications for platform designers, community moderators, and policymakers working to develop more effective digital spaces that can enhance the benefits of online interaction while mitigating its potential harms.