Personalized Algorithms and the Virtue of Learning Things the Hard Way

Abstract

Personalized recommendation systems are now an integral part of the digital ecosystem. However, consumers' increased dependence on these personalized algorithms has heightened concerns among consumer protection advocates and regulators. Past studies have documented various threats personalization algorithms pose to different aspects of consumer welfare, through violating consumer privacy, unfair allocation of resources, or creating filter bubbles that can lead to increased political polarization. In this work, we bring a consumer learning perspective to this problem and examine whether personalized recommendation systems hinder consumers' independent decision-making ability, an important construct given the growing fear of adversarial AI. We develop a utility framework where consumers learn their preference parameters through experience to examine the effect of personalized algorithms on the learning process. We establish regret bounds for different types of consumers based on their dependence on the personalized algorithm. We then run a series of calibrated simulations and show that although personalized algorithms increase consumer welfare for consumers who rely more on personalized recommendations by offering better recommendations, these consumers do not sufficiently learn their own preference parameters and make worse decisions in the absence of recommendation systems. Inspired by the trade-off between consumer learning and welfare, we introduce the notion of counterfactual regret, which is the regret incurred by the consumer when the personalized algorithm is unavailable. Finally, we examine a variety of consumer protection policies that aim to find a balance between these two outcomes and find policies that can achieve good welfare and learning outcomes.

Keywords: personalized algorithms, recommendation system, consumer learning, consumer protection, linear bandits, reinforcement learning

JEL Classification: L51, D83