Technical Program

Paper Detail

Paper Title An Information-Theoretic Look at Fair Machine Learning
Paper IdentifierFR3.R1.4
Authors Flavio P. Calmon, Harvard University, United States
Session Information Privacy and Fairness
Location Le Théatre (Parterre), Level -1
Session Time Friday, 12 July, 14:30 - 16:10
Presentation Time Friday, 12 July, 15:30 - 15:50
Abstract Machine learning (ML) algorithms are increasingly used in applications of significant social consequence. These algorithms are at risk of inheriting and, ultimately, reinforcing social biases and discrimination patterns present in the training data, which may reflect discrimination patterns existent in society at large. Automated methods for discovering and controlling discrimination in ML may inherently trade-off between fairness and accuracy. In this talk, we present a few information-theoretic results that build on rate-distortion theory and estimation theory to characterize the fundamental accuracy cost of ensuring fairness in ML. We start by characterizing the limits of data pre-processing for discrimination control by extending wellknown formulations found in rate-distortion theory. We then delineate a few fundamental limits of non-discrimination in classification by studying fairness-constrained maximum a posteriori (MAP) estimators. This approach informs a series of benchmarks based on artificial data that can be used to systematically compare and evaluate the growing number of fairness intervention mechanisms available in the ML literature. This work was done in collaboration with Hao Wang, Shahab Asoodeh, and Wael Alghamdi.