Learning with Feedback Loops
Many learning processes have the feature that realizations of the data affect the data generating process itself.
Abstract
Many learning processes have the feature that realizations of the data affect the data generating process itself. I investigate the impact of this phenomena—which I term “feedback loops”—on learning processes mediated by recommendation systems. In particular, I characterize the learning outcomes when the platform is naïve in the sense that it fails to account for feedback in the data. Naïve platforms exhibit failures of asymptotic learning, even when agents possess arbitrarily precise information. This results in persistent errors in recommendations, which can influence the actions of a large number of agents in equilibrium. Conversely, platforms which correctly account for feedback guarantee asymptotic learning and correct herding. My findings emphasize the critical need for recommendation systems to account for feedback in the data.
The paper is currently undergoing revisions.