Distributionally Robust Optimization with Probabilistic Group

Modern machine learning models may be susceptible to spurious correlations that hold on average but not for the atypical group of samples. To address problem, previous approaches minimize empirical worst-group risk. Despite promise, they often assume each sample belongs one and only group, which does allow expressing uncertainty in labeling. In this paper, we propose a novel framework PG-DRO, explores idea probabilistic membership distributionally robust optimization. Key our framework, consider soft instead hard annotations. The probabilities can flexibly generated using either supervised or zero-shot approaches. Our accommodates samples with ambiguity, offering stronger flexibility generality than prior art. We comprehensively evaluate PG-DRO both image classification natural language processing benchmarks, establishing superior performance.