Data-driven decision making systems are increasingly used to guide important decisions ranging from loan approvals to school admissions to prison sentencing guidelines. Existing methods for providing these decisions are often exclusively designed to maximize a single criterion (e.g., accuracy, utility). For socially-impactful applications, such an approach ignores most of the other important decision considerations. Moreover, most of the current methods assume that decisions are made once or are independent of previous similar decisions, that all the information necessary for the decision is available, and that a single criterion of fairness can characterize the decision. Not surprisingly, these assumptions do not hold in the real world. This project investigates education, public health, and urban revitalization decision-making applications in the City of Chicago, as well as environmental policy applications in collaboration with Wild Me, an AI for conservation non-profit. The project aims to develop a fair machine learning approach that takes into account deficiency of information, dynamic decision making, and disagreement about a single fairness criterion.
This project aims to develop a fair machine learning approach based on robust estimation to address three forms of complexity common in application domains: deficiency of information, dynamic decisions, and disagreement about a single fairness criterion. It approaches information deficiency by generalizing the notion of proxies and learning latent group membership structure as a transfer learning task. It flexibly extends the notion of fairness to dynamic settings with repeated interactions by defining it over states of the individuals rather than decision maker actions. Finally, it addresses fairness disagreements by rationalizing the fairness criteria of observed decisions and balancing different criteria.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
