Mitigating Spurious Correlation via Distributionally Robust Learning with Hierarchical Ambiguity Sets

2025-10-03

AI Summary

Conventional supervised learning is vulnerable to spurious correlations under distributional shifts, particularly when minority-group samples are scarce. Existing methods—such as Group Distributionally Robust Optimization (Group DRO)—only mitigate inter-group shifts and fail to address intra-group distributional changes. Method: We propose Hierarchical Robust Learning (HRL), the first framework to model dual uncertainty—both inter-group and intra-group—via hierarchical ambiguity sets. We extend Group DRO into a bilevel worst-case risk minimization formulation that jointly optimizes robustness across and within groups. Additionally, we construct the first benchmark dataset explicitly designed to simulate intra-group shifts in minority subpopulations. Contribution/Results: Experiments demonstrate that HRL significantly outperforms state-of-the-art robust learning methods on both standard benchmarks and our new intra-group shift benchmark. It maintains stable performance under minority-group distribution shifts, thereby advancing multi-level distributional robustness.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Intelligent Robots: Learning & Optimization for ROB
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

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.

Problem

• Addresses vulnerability to spurious correlations under distribution shifts
• Mitigates intra-group uncertainties in minority groups with limited samples
• Provides robustness to hierarchical distribution shifts at multiple levels

Innovation

• Hierarchical extension of Group DRO method
• Addresses inter-group and intra-group uncertainties
• Introduces new benchmark for minority shifts