Theoretical studies of opinion formation and evolution in social networks often focus on convergence to a set of steady-state opinions, namely asymptotic learning (with or without consensus). This is often motivated by the desire to seek an 'equilibrium' according to various definitions from statistical physics, control engineering, or economics. In many real social settings, however, it is observed empirically that opinions do not converge to a steady state; instead, they fluctuate indefinitely. This interdisciplinary workshop has three goals: (i) to introduce attendees to fundamental theoretical tools based on Bayesian inference that are suitable for modelling opinions and their evolution; (ii) to highlight some counter-intuitive yet realistic social phenomena that emerge when applying these tools; and (iii) to bring together practitioners from different knowledge domains (e.g. media studies, political science, education, artificial intelligence, social sciences, complex systems, and network sciences), who aspire to apply the tools to real-life systems. Specifically, we begin with an introduction to Bayesian statistics and belief propagation over networks. This enables us to learn the underlying tools required for modelling and analysis of opinion evolution across social networks. At this stage, we will also review some results on asymptotic learning on social networks facilitated by Bayesian inference. We then delve into a new model of opinion formation and evolution by enmeshing Bayesian learning and peer interactions. As an illustrative example, we consider a scenario where networked agents form beliefs about the political bias of a media organisation through consumption of media products, and peer pressure from political allies and opponents. To capture the multi-modal nature of opinions (individuals can hold contradictory beliefs with different levels of certainty), we model the agents' beliefs as probability distribution functions. In certain network structures, numerical simulations reveal counter-intuitive predictions, such as wrong conclusions being reached quicker with more certainty, turbulent non-convergence (some agents cannot “make up their mind” and vacillate in their beliefs), and intermittency (agents' beliefs flip between stable eras, where their beliefs do not vary over many time steps, and turbulent eras, where their beliefs fluctuate from one time step to the next). We will also consider belief disruption by partisans, i.e. stubborn agents who do not change their beliefs. If time permits, attendees will receive practical, hands-on instruction in coding the methods covered during the workshop.

Workshop Length: 3 hours.

Maximum Attendees: 30.