Pairwise and Higher-order Mixing Patterns in Online Social Networks

Online social platforms are central arenas for contemporary public discourse. Yet, a growing body of research shows that the interplay between human cognitive biases and algorithmic mediation can foster concerning dynamics such as opinion polarization, radicalization, and the formation of echo chambers. These phenomena arise from the interaction patterns that determine who encounters whom (or what), and under what conditions. In this context, user mixing (i.e., how users with similar/different attributes cluster) constitutes a fundamental mechanism linking individual behavior to collective outcomes. Existing work on mixing, however, has focused predominantly on static, pairwise network representations. This leaves open important questions about how mixing patterns persist and transform when interactions are time-varying and/or organized in higher-order structures. Addressing this gap, by studying mixing across multiple structural and temporal scales, is key both for developing explanatory theories of online social dynamics and for informing the design of more inclusive and algorithmically fair platforms. \\ \ \\ Building on Part I, which provides social and technical background, this thesis sets three complementary objectives. The first objective is to characterize how pairwise mixing between classes of users shapes online debates. To this end, Part II focuses on dyadic interactions in large-scale social media data. In a Covid-19 vaccination debate on Italian Twitter, the thesis quantifies how authority accounts influence the opinions of their audiences, capturing both cross-stance exposure and opinion shifts. In a second case study on climate change discussions, we study acrophily, a peer-level mixing pattern where users preferentially interact with ideologically extreme in-group members, and show statistically significant cross-class differences. The second objective is to extend the analysis of mixing patterns from dyads to groups and higher-order interaction structures. To this end, Part III introduces Attributed Stream Hypergraphs (ASH), a temporal model of node-attributed higher-order interactions. ASH supports a family of structural and mixing-related measures that jointly capture how users move through, and are exposed to, heterogeneous group contexts over time. Finally, it proposes a general schema to extend classical graph-based mixing indices to hypergraphs and introduces a novel local random-walk-based measure that quantifies how strongly nodes and interaction contexts are trapped within same-class neighborhoods. The third objective is to mitigate data-access limitations and to create methodological tools that support the long-term study of mixing and influence in online environments. Part IV addresses this by (i) constructing and releasing a large-scale high-coverage dataset that combines social ties, content, and algorithmic activity, and (ii) developing LLM-driven simulation environments for opinion dynamics and social media ``digital twins''. These simulated settings make it possible to study the interplay between algorithmic and human behaviors under controlled and reproducible conditions that are increasingly hard to obtain from real platforms alone. Taken together, the contributions of this thesis provide (i) conceptual frameworks for thinking about mixing patterns beyond static dyads, (ii) methodological tools for quantifying temporal, pairwise, and higher-order mixing in both observational and simulated data, and (iii) empirical insights into how user behavior and platform design interact to produce homophilic and heterophilic patterns in online social systems. Most importantly, we show that temporal and higher-order perspectives are non-trivial to obtain, yet they reveal mixing and exposure structures that remain hidden in static pairwise representations.