From hostility to hyperlinks: mining social networks with heterogenous ties

Thesis

Social networks are a fundamental tool for understanding emergent behaviour in human society, providing a mathematical framework that emphasizes the importance of interactions between the individuals in the network. While traditional social network models consider all social ties as uniform, either an edge exists or it does not, human nature is more complex and individuals can be linked by relationships that differ in nature, intensity, or sentiment. This tie-level complexity can be represented using more complex network models, including signed, weighted and multiplex networks, where edge-level attributes delineate between the types of interactions. A growing body of literature is devoted to developing methods for extracting information from such heterogenous networks, from probing the latent structure to investigating dynamical processes occurring overtop of them.

In this thesis, we focus on ties that vary in sentiment, using signed networks in which edges carry positive or negative weights, representing cooperative or antagonistic relationships, and ties that vary in nature, using weighted and multiplex network models. We present models and empirical studies that adapt traditional network science methods to extract information, detect multi-scale structure and characterize dynamical processes, to the heterogeneous network context. Our contributions here are four-fold. First, we present a signed network embedding algorithm capable of recovering continuous node attributes. Next, we develop a signed network random walk, which effectively captures macro-scale graph structure. Third, we conduct an empirical study of peer-to-peer music recommendations, using a model of social contagion that leverages the information contained in heterogeneous ties. Our final work draws upon the learnings from the previous chapters, presenting a study of political polarization on a social media platform, represented as a signed, weighted and multiplex network. Overall, this thesis presents methodological and empirical advances, which demonstrate the advantage of maintaining tie-level complexity in mining social networks.

Authors: Babul, SAV

• DOI: 10.5287/ora-b752ywj8x
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: social dynamics; dynamical systems; complex systems; network science; social networks
• Subjects: Social sciences--Network analysis; Coupled problems (Complex systems)