CAREER: Robust Policy Learning for Safe and Reliable Algorithmic Decision Making from Observational Data in Sensitive Applications

Cornell University, Ithaca NY

Investigators

Abstract

Some of the most impactful applications of machine learning are not just about prediction but are rather about taking the right action directed at the right target at the right time. Actions, unlike predictions, have consequences and so, in seeking to take the right action, one must seek to understand its causal effect. This project deals with the problem of extracting causal-effect-maximizing personalized decision rules from observational data in sensitive applications. Observational data, which have become plentiful in domains such as medicine and civics, lack experimental manipulation so that isolated causal effects are obscured by complex selection processes, a phenomenon known as confounding, and such data are also otherwise messy, noisy, biased, and often missing. Despite the promise of rich and plentiful observational data, current approaches cannot handle the unique challenges it poses and can lead to unreliable, unsafe, and unfair decision making unfit for sensitive applications. The goal of this project is to create a comprehensive framework of rigorous theory and robust methodology to address this gap and enable trustworthy decision-making systems trained on observational data. The research and education plans are integrated through student advising and curriculum development and include targeted outreach efforts to broaden participation of underrepresented groups. The research will proceed along three primary directions. The first is to develop methods and theory for algorithmic decision making in the presence of unobserved confounders. Whereas existing approaches tenuously rely on various unverifiable assumptions that ensure point-identification of causal effects, the project will develop robust learning methods that produce policies with certificates of safety and/or improvement backed by theoretical guarantees that do not rely on exact identification. The second is to develop methods and theory for robust and optimal weighting for policy learning to address issues of stability, limited overlap, time-to-event data, and noisy and missing observations. The third is to investigate algorithmic fairness of decision policies trained from observational data. The project will develop characterizations of fairness in settings where disparity metrics cannot be point-identified from data, whether due to biased selection or missing attributes and labels, as well as methods that can audit and enforce fairness in such settings. Through these thrusts, the project will both advance knowledge at the intersection of machine learning, causal inference, and optimization as well as broaden the scope of their integration with new problem domains. 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.