Link recommendation algorithms and opinion polarization

by Fernando Santos

Online social media platforms are, today, spaces where political opinions are formed, reinforced, and

confronted. These platforms are also environments where humans co-exist with algorithms. Hence,

understanding opinion dynamics requires comprehending the interrelated subtleties of human

decision-making and the outcomes of automated decisions. As high levels of political polarization

raise concerns in different parts of the world, we should ask: What is the impact of social network

algorithms in this process? Can such algorithms be used as an intervention mechanism to control

polarization?

In a recent paper published in the Proceedings of the National Academy of Sciences (special feature on Dynamics of Opinion Polarization) we provide a complex adaptive systems perspective on the effects of link recommendation algorithms in opinion polarization. These algorithms — also called user, contact, or people recommender systems— are used to recommend new connections to users. We show that link recommendation algorithms that suggest connections to be established between users already sharing many “friends” — as it is likely to be the case in current applications — lead tight-knit communities to emerge. This exacerbates polarization as isolated communities of users can perpetuate diverging opinions and individuals are less likely to be exposed to a diverse pool of viewpoints.

To study the co-evolving dynamics of link recommendations and opinion formation, we develop a new computational model to simulate the evolution of a social network alongside influence between individuals with different opinions. This model allows us to test how tuning the dependence of link recommendations on the number of common friends might impact the social network topology and, in turn, impact opinion polarization.

Our study is relevant for three key reasons: first, it sheds light on the impacts of social-network algorithms on dynamics of polarization and consensus; second, it suggests that small modifications in how recommendations are made can significantly impact information flows and opinion polarization; finally, and more broadly, it reveals that understanding the impacts of algorithms in our society requires framing their effects in the context of complex adaptive systems where individuals and automated decisions (co-)adapt to each other over time.

Overall, our study stresses the impacts of social-network algorithms and unveils avenues to control dynamics of radicalization and polarization on online social networks. We hope this inspires individuals, algorithm designers and policymakers to recognize the potential impacts of social network algorithms in long-term opinion dynamics and political polarization. While our article explores a process of link formation based on common friends, we invite new theoretical and empirical analysis that investigate how more complex link recommendations might help to establish connections that are both useful for users and can contribute to curbing polarization on social media.

More info at: https://www.pnas.org/content/118/50/e2102141118

And the full Special Feature at: https://www.pnas.org/cc/dynamics-of-political-polarization-special-feature