A0418
Title: Empirical Likelihood for fair classification
Authors: Pangpang Liu - Purdue University (United States)
Yichuan Zhao - Georgia State University (United States) [presenting]
Abstract: Machine learning algorithms are commonly being deployed in decision-making systems that have a direct impact on human lives. However, if these algorithms are trained solely to minimize training/test errors, they may inadvertently discriminate against individuals based on their sensitive attributes, such as gender, race or age. Recently, algorithms that ensure fairness have been developed in the machine learning community. Fairness criteria are applied by these algorithms to measure fairness, but they often use the point estimate to assess fairness and fail to consider the uncertainty of the sample fairness criterion once the algorithms are deployed. It is suggested that fairness should be assessed by taking uncertainty into account. The covariance is used as a proxy for fairness and develops the confidence region of the covariance vector using empirical likelihood. The confidence region-based fairness constraints for classification take uncertainty into consideration during fairness assessment. The proposed confidence region can be used to test fairness and impose fairness constraints using the significance level as a tool to balance accuracy and fairness. Simulation studies show that the method exactly covers the target Type I error rate and effectively balances the trade-off between accuracy and fairness. Finally, data analysis is conducted to demonstrate the effectiveness of the method.
