Assistant Professor Chicheng Zhang received an Outstanding Paper Runner Up Award at the recent International Conference on Machine Learning (ICML 2022), one of the top conferences in machine learning. The paper, co-authored by Dr. Zhang and Tom Yan (at Carnegie Mellon University) is titled “Active Fairness Auditing”. Among over 1200 accepted papers at ICML this year, 10 were selected as Outstanding Papers, and 5 were selected as Outstanding Paper Runner Up.
The paper is motivated by the recent observation that the fast adoption of machine learning (ML) by companies across industries (e.g. for hiring, loan approval) poses significant regulatory challenges. It tackles the following question: how can regulatory bodies audit the fairness of these ML models in a minimally-intrusive manner, in the sense that it makes only a few query access to the model predictions? The paper initiates the study of query-efficient fairness estimation algorithms, with a focus on estimating the demographic parity of ML models (i.e. the difference of the positive prediction rates of a machine learning model across two demographic groups). Specifically, the paper investigates the fundamental query and computational complexities of this problem, providing: (1) an optimal deterministic algorithm; (2) an efficient randomized algorithm with competitive guarantees; (3) several fundamental information-theoretic limits for fairness estimation.
Dr. Zhang joined the University of Arizona Department of Computer Science in August 2019. You can watch Dr. Zhang’s ICML 2022 presentation recording here.