Chara Podimata, Harvard University
As machine learning algorithms are increasingly being deployed for consequential decision making (e.g., loan approvals, college admissions, probation decisions etc.) humans are trying to strategically change the data they feed to these algorithms in an effort to obtain better decisions for themselves. If the deployed algorithms do not take these incentives into account they risk creating policy decisions that are incompatible with the original policy’s goal.
In this talk, I will give an overview of my work on Incentive-Aware Machine Learning for Decision Making, which studies the effects of strategic behavior both to institutions and society as a whole and proposes ways to robustify machine learning algorithms to strategic individuals. I will first explain the goals of the different stakeholders (institution, individual, society) in these settings in a unified way and show the various settings I have worked on that belong in the incentive-aware machine learning area such as incentive-compatible algorithms for linear regression and online prediction with expert advice, strategic classification, learning in auctions, and dynamic pricing. I will conclude by looking at the problem from a societal lens and discuss the tension that arises between having decision-making algorithms that are fully transparent and incentive-aware.
Bio: Chara is a final year PhD student at Harvard, where she is advised by Yiling Chen. Her research is supported by a Microsoft Dissertation Grant and a Siebel Scholarship. During her PhD, she interned twice for MSR NYC (mentored by Jennifer Wortman Vaughan and Aleksandrs Slivkins) and once for Google Research NYC (mentored by Renato Paes Leme). She has given tutorials related to strategic learning at EC20 and FAccT21. Outside of research, she spends her time adventuring with her pup, Terra.