Class-attribute Priors: Adapting Optimization to Heterogeneity and Fairness Objective

Xuechen Zhang, Mingchen Li, Vala Vakilian, Jiasi Chen, Christos Thrampoulidis, Samet Oymak

Keywords: fairness, personalization, imbalance, bilevel optimization

Abstract: Modern classification problems exhibit heterogeneities across individual classes: Each class may have unique attributes, such as sample size, label quality, or predictability (easy vs difficult), and variable importance at test-time. Without care, these heterogeneities impede the learning process, most notably, when optimizing fairness objectives. We propose an effective and general method to personalize the optimization strategy of individual classes so that optimization better adapts to heterogeneities. Concretely, class-attribute priors (CAP) is a meta-strategy which proposes a class-specific strategy based on its attributes. This meta approach leads to substantial improvements over naive approach of assigning separate hyperparameters for each class. We instantiate CAP for loss function design and posthoc logit adjustment, with an emphasis on label-imbalanced problems. We show that CAP is competitive with prior art and its flexibility unlocks noticeable improvements for fairness objectives beyond balanced accuracy. Finally, we evaluate CAP on problems with label noise as well as weighted test objectives to showcase how CAP can synergistically leverage different class attributes.

Submission Checklist: Yes

Broader Impact Statement: Yes

Paper Availability And License: Yes

Code Of Conduct: Yes

Reviewers: Yes

CPU Hours: 0

GPU Hours: 0

TPU Hours: 0