Skip to content

Explainer: What Is Distributionally Robust Optimization (DRO) for Fairness? #544

Description

@yakew7

Add explainers/distributionally-robust-optimization.md.

None of this repo's 56 explainers cover Distributionally Robust Optimization (DRO) as a fairness in-processing technique. Mitigation Strategies already covers this repo's own S3 in-processing strategy (Fairlearn's ExponentiatedGradient, a constraint-based approach: optimize accuracy subject to a fairness constraint like demographic parity). DRO is a genuinely different in-processing philosophy worth its own explainer: instead of constraining to a specific fairness metric, train the model to minimize its worst-case loss over an uncertainty set of possible subgroup distributions - without needing to know group membership as a hard constraint, and without targeting one specific parity metric at all.

Suggested structure (match mitigation-strategies.md's in-processing section, as its own dedicated explainer): one-sentence definition -> why it matters (a constraint-based method like ExponentiatedGradient needs the fairness metric decided up front and group membership at training time; DRO instead hedges against the worst subgroup the training data might be under-representing, which can help even when the actual harmful subgroup isn't the one you thought to check) -> core concept (the min-max objective: minimize the maximum expected loss over a distribution-shift ball around the empirical training distribution, contrasted directly against ExponentiatedGradient's constrained-optimization objective already described in mitigation-strategies.md) -> a concrete example - implement a simple group-DRO training loop (upweighting whichever group currently has the highest loss during training, a standard simplified DRO approach) against one of this repo's real audits, and compare the resulting per-group performance against this repo's own frozen S0/S3 baseline numbers, with real code and real (not invented) results -> detection/implementation code -> limitations (DRO's uncertainty set size is a free parameter - too large and the model over-hedges against implausible worst cases, hurting average-case accuracy for no real benefit; still needs group labels during training even though it doesn't optimize a specific parity metric) -> related concepts (mitigation-strategies, fairness-accuracy-tradeoff, class-imbalance) and projects.

Related: mitigation-strategies, fairness-accuracy-tradeoff, class-imbalance.

Key citation to verify and use: Sagawa, Koh, Hashimoto, Liang (2020), "Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization" (ICLR) - confirm the exact venue/year/finding yourself before citing, don't trust this issue's paraphrase.

See CONTRIBUTING.md: add explainers/<slug>.md plus a one-line entry in assets/explainers-data.json; the build script generates the page. Per CLAUDE.md, the paper freeze is lifted - if you quote a Fair Code benchmark result, use the current numbers in results/ (or paper/results-frozen/ for the earlier reference snapshot) and say which one.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    documentationImprovements or additions to documentationhelp wantedExtra attention is needed

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions