Distributionally Robust Losses for Latent Covariate Mixtures

Reliable Machine Learning via Structured Distributionally Robust Optimization Data sets used to train machine learning (ML) models often suffer from sampling biases and underrepresent marginalized groups. Standard are trained optimize average performance perform poorly on tail subpopulations. In “Distributionally Losses for Latent Covariate Mixtures,” John Duchi, Tatsunori Hashimoto, Hongseok Namkoong formulate a DRO approach training ML uniformly well over They design worst case optimization procedure structured distribution shifts salient in predictive applications: (a subset of) covariates. The authors propose convex that controls subpopulation provide finite-sample (nonparametric) convergence guarantees. Empirically, they demonstrate their lexical similarity, wine quality, recidivism prediction tasks observe significantly improved across unseen