Multiaccurate Proxies for Downstream Fairness

Watch a research presentation from the 2022 Symposium on Foundations of Responsible Computing where Emily Diana from the University of Pennsylvania explores how to train machine learning models to maintain demographic fairness without direct access to sensitive features during training. Learn about a novel fairness pipeline approach that uses an "upstream" learner with access to sensitive features to create proxy models, enabling "downstream" learners to develop fair models despite lacking direct access to demographic data. Discover how multiaccuracy constraints can be effectively used to ensure fairness, with demonstrations of sample- and oracle-efficient algorithms, generalization bounds, and experimental results showing that multiaccuracy requirements are more achievable than strict classification accuracy when predicting sensitive attributes.