Incentive-Aware Machine Learning for Decision Making

Abstract

Machine Learning algorithms are increasingly being deployed in consequential decision-making for people's lives. These decisions affect widely different aspects of our lives; e.g., Machine Learning algorithms decide what content to serve us online (thus guiding our purchasing behavior and overall beliefs) or whether we are creditworthy enough for a loan. In light of how consequential these algorithmic decisions are, people have been documented to ``strategize'' with the data that they feed to the Machine Learning algorithms hoping to obtain better outcomes or decisions. This dissertation focuses on such decision-making settings where people are incentivized to react to the algorithmic decisions made. Incentive-aware Machine Learning for decision-making creates a new ecosystem with three key stakeholders: the institutions deploying the algorithms, the individuals who are being impacted by the algorithms, and society as a whole. It is natural to expect that oftentimes there is tension between the goals of these stakeholders. In this dissertation, we study the tension that arises from each of the stakeholders' perspectives. We start by thinking about the problem of algorithmic decision making from the institution's perspective. Drawing intuition from the literature on Algorithmic Game Theory, we build new machine learning algorithms that align the incentives of the individuals with these of the institution (incentive compatibility), while not sacrificing too much accuracy. Incentive-compatibility is the holy grail of incentive alignment and is sometimes unattainable. To alleviate this, we propose Machine Learning algorithms that learn to adapt to the incentives of the agents that they face, hence also ensuring a form of robustness. Shifting gears, we focus on the societal implications of information discrepancy regarding the deployed machine learning algorithms to different subpopulations. Finally, we study incentive-aware learning in settings where there are misspecifications regarding the behavioral model of the agents and incentive-aware learning in settings where the agents are non-myopic (and sometimes, they are even learners themselves). This dissertation presents foundational theoretical advancements that range from algorithms for ubiquitous tasks such as linear regression and classification, prediction with expert advice, and Lipschitz bandits to theoretical results tailored to specific application domains such as credit scoring, forecasting, pricing, and auctions.