Microfoundations of ML: Designing for Strategic Human Behavior

As machine learning algorithms increasingly automate high-stakes decision-making in areas like hiring, lending, and judicial sentencing, individuals often respond strategically to influence their outcomes. These responses create 'moving targets' for ML algorithms due to the reactive nature of human-generated data. In this talk, I will present our group's recent works on managing this challenge, focusing on the impacts of two forms of biased/incomplete information.   

First, I will present a prospect-theoretic model to capture the impacts of humans' cognitive biases when responding strategically to algorithmic systems. I will complement our analytical findings based on this model with supporting findings from user studies. Second, we will consider the evolving and uncertain nature of agents' strategic response costs over time. To capture this, we model and analyze algorithm selection as a two-stage robust optimization problem with decision-dependent uncertainty. Our findings here highlight the role of planning in both reducing uncertainty and reducing undesired strategic behavior.