Robustness to missing validation subgroups

Establish robustness guarantees for POTER when one or more validation subgroups are entirely missing from the group-annotated reference set.

Background

POTER constructs its reference distribution from a representative, class-labeled validation set with subgroup annotations. The method can be evaluated with reduced reference size, noisy reference labels, or without group annotations under some conditions, but its standard formulation assumes that the validation data provide coverage of the relevant subgroups.

The paper explicitly identifies the case in which validation subgroups are entirely absent as unresolved. This is important because missing subgroups may prevent the reference distribution from representing the target group composition and may undermine the optimal-transport sample weights used for robust empirical risk minimization.

References

One limitation is that POTER still depends on a representative, class-labeled validation set. Appendix~\ref{sec:appendix-validation-robustness} examines reduced and noisy references, while robustness to entirely missing validation subgroups remains an open challenge.

— Optimal Transport Reweighting for Robust Learning under Spurious Correlations and Label Noise  (2610.01028 - Jo et al., 1 Oct 2026) in Conclusion