A complex systems perspective of news recommender systems: Guiding emergent outcomes with feedback models

Algorithms are increasingly making decisions regarding what news articles should be shown to online users. In recent times, unhealthy outcomes from these systems have been highlighted including their vulnerability to amplifying small differences and offering less choice to readers. In this paper we present and study a new class of feedback models that exhibit a variety of self-organizing behaviors. In addition to showing important emergent properties, our model generalizes the popular "top-N news recommender systems" in a manner that provides media managers a mechanism to guide the emergent outcomes to mitigate potentially unhealthy outcomes driven by the self-organizing dynamics. We use complex adaptive systems framework to model the popularity evolution of news articles. In particular, we use agent-based simulation to model a reader's behavior at the microscopic level and study the impact of various simulation hyperparameters on overall emergent phenomena. This simulation exercise enables us to show how the feedback model can be used as an alternative recommender to conventional top-N systems. Finally, we present a design framework for multi-objective evolutionary optimization that enables recommendation systems to co-evolve with the changing online news readership landscape.