Closed-Loop Opinion Formation

Abstract

When information sources are moderated by recommender systems, so-called "filter bubbles" may restrict the diversity of content made available to users, potentially affecting their opinions. User opinions may in turn affect the output of recommender systems. In this work we ask how the dynamical system defined by user and recommender systems behaves, as each element evolves in time. In particular, we look at whether the use of recommender system can affect user experience and user opinions in a systematic way. We define and analyze three metrics to understand those effects - intensity, simplification, and divergence - and we explore both link-based and ratings-based recommender systems. Our results suggest that previous studies of this problem have been too simplistic, and that user opinions can evolve in complex ways under the influence of personalized information sources.