Distributionally Robust Losses for Latent Covariate Mixtures

• 成果类型: Article
• 署名作者: Duchi, John; Hashimoto, Tatsunori; Namkoong, Hongseok
• 署名单位: Stanford University; Stanford University; Stanford University; Columbia University
• 刊物名称: OPERATIONS RESEARCH
• ISSN/ISSBN: 0030-364X
• DOI: 10.1287/opre.2022.2363
• 发表日期: 2023
• 页码: 649-664
• 关键词: inference; prediction; shift

摘要

While modern large-scale data sets often consist of heterogeneous subpopulations-for example, multiple demographic groups or multiple text corpora-the standard practice of minimizing average loss fails to guarantee uniformly low losses across all sub-populations. We propose a convex procedure that controls the worst case performance over all subpopulations of a given size. Our procedure comes with finite-sample (nonparametric) convergence guarantees on the worst-off subpopulation. Empirically, we observe on lexical similarity, wine quality, and recidivism prediction tasks that our worst case procedure learnsmodels that do well against unseen subpopulations.