Job Done? Future Modeling Challenges After 20 Years of Work on Bounded-Confidence Models

Since the first publication of the bounded-confidence models 20 years ago, hundreds of articles studying this class of social-influence models have been written. Bounded-confidence models proposed an intriguing solution to a pervasive research puzzle and have helped unveil and explain intriguing phenomena. Here, we reflect about remaining research problems and future modeling challenges, arguing that there remain counter-intuitive model implications to be understood. To illustrate that there remain uncovered model challenges, we extend the bounded-confidence model. We assume assimilative influence when agents connected by positive relationships hold sufficiently similar opinions, adopting the core assumption of the bounded-confidence models. We combine this with another influential modeling approach, the notion that if agents connected by negative social relationship disagree too much, opinion differences increase due to repulsive influence. This allows us to vary the relative strength of assimilation and repulsion in the influence dynamics, also allowing for the possibility that neither occurs in a particular interaction. Simulation experiments reveal three surprising findings: Counter the intuition that stronger assimilation decreases opinion diversity, we show that in the presence of repulsion, intensifying the strength of assimilation can actually generate more opinion bipolarization. Second, we show that if repulsion becomes weaker this may still result in more bipolarization. Third, it turns out that more negative social relationships between or within subgroups can result in less bipolarization. We demonstrate these effects in very simple and highly stylized settings, in order to show that intuition fails to capture the complexity arising from the interplay of assimilative and repulsive influence even in these simple settings. We discuss implications of our findings for the ongoing debate about societal conditions fostering bipolarization, including in particular the design of personalized online social networks. Further, we address how our results may inform future work comparing and integrating alternative models of social-influence dynamics.