Robust Fairness Under Covariate Shift

Making predictions that are fair with regard to protected attributes (race, gender, age, etc.) has become an important requirement for classification algorithms. Existing techniques derive a model from sampled labeled data relying on the assumption training and testing identically independently drawn (iid) same distribution. In practice, distribution shift can does occur between datasets as characteristics of individuals interacting machine learning system change. We investigate fairness under covariate shift, relaxation iid in which inputs or covariates change while conditional label remains same. seek decisions these assumptions target unknown labels. propose approach obtains predictor is robust worst-case performance satisfying requirements matching statistical properties source data. demonstrate benefits our benchmark prediction tasks.