Mitigating Spurious Correlation via Distributionally Robust Learning with Hierarchical Ambiguity Sets

Conventional supervised learning methods are often vulnerable to spurious correlations, particularly under distribution shifts in test data. To address this issue, several approaches, most notably Group DRO, have been developed. While these methods are highly robust to subpopulation or group shifts, they remain vulnerable to intra-group distributional shifts, which frequently occur in minority groups with limited samples. We propose a hierarchical extension of Group DRO that addresses both inter-group and intra-group uncertainties, providing robustness to distribution shifts at multiple levels. We also introduce new benchmark settings that simulate realistic minority group distribution shifts-an important yet previously underexplored challenge in spurious correlation research. Our method demonstrates strong robustness under these conditions-where existing robust learning methods consistently fail-while also achieving superior performance on standard benchmarks. These results highlight the importance of broadening the ambiguity set to better capture both inter-group and intra-group distributional uncertainties.

💡 Research Summary

This paper addresses a critical weakness of existing robust learning methods for spurious correlations, namely their vulnerability to intra‑group distribution shifts that commonly affect minority groups with few training samples. While Group Distributionally Robust Optimization (Group DRO) protects against changes in group proportions (inter‑group shifts), it assumes that each group’s conditional distribution is accurately captured by the limited training data. In practice, minority groups often exhibit substantial internal variability, leading to severe performance degradation when the test distribution deviates from the training estimate for that group.

To remedy this, the authors propose a hierarchical ambiguity set that simultaneously models uncertainty over group mixing proportions and over the conditional distributions within each group. Formally, the training distribution is expressed as a mixture (P = \sum_{g=1}^m \alpha_g P_g). The new ambiguity set is defined as
\