Social networks and polarization

Depending on how it is used, structured and harnessed, social media can be a breeding ground for social division and extremist ideologies or a way to foster social cohesion through the sharing of diverse viewpoints. We rely on social learning models to understand why some societies do not aggregate information as accurately, quickly or cohesively as others. These models allow us to see how the way that influence is distributed across a society affects outcomes. However, many of the standard models are not suitable for analysing domains where information is abundant but contentious, as they incorporate strong consensus-building assumptions and are therefore unable to result in long-term polarization of views.

I build a class of DeGroot-style social learning models that can result in polarization. I use these to explore which features of network-based information sharing support consensus-building and which enable division. I do this by allowing the distance between views to affect how much attention is paid to them, with a bias towards consuming pro-attitudinal information - i.e. the act of listening more to people whose views align more closely with one's own. I show that long-term polarization hinges both on how opposing views are processed by recipients as well as on the network's structure. Polarization is minimised when (1) agents continue to listen at all, however slightly, to a wide range of opposing views, but only when recipients experience no `backlash' inducing reaction (2) moderates have greater relative influence than extremists and (3) the level of assortativity (homophily) is low. The models highlight potential impacts of policies often thought to reduce polarization, such as deplatforming or reducing influence of extremists on social media, facilitating inter-group dialogue in conflictual settings, providing exposure to counter-attitudinal views or using widely trusted actors rather than lesser-known experts to convey contentious information.