Fairness by Design: Efficient Fair Ensembles for Low-Data Classification

ICLR 2026 Conference Submission18133 Authors

Keywords: Algorithmic fairness, Group fairness, Minimum recall, Equal opportunity, Ensemble learning, Low-data learning, Medical imaging, Theoretical guarantees

TL;DR: We introduce fair deep ensembles that work well with extremely low-data groups with theoretical guarantees.

Abstract:

We address the problem of fair classification in settings where data is scarce and unbalanced across demographic groups. Such low-data regimes are common in domains like medical imaging and hate speech. Our proposed method mitigates these biases by training efficient ensembles of fair classifiers on different data partitions. Aggregating predictions across ensemble members, each trained to satisfy fairness constraints, yields more consistent outcomes and stronger fairness-accuracy trade-offs than existing methods across multiple challenging medical imaging datasets, as well as on hate speech detection.

To support these findings, we provide theoretical guarantees: we prove when our fair ensembles improve performance and how much data is needed to observe these gains with statistical significance. These results extend the literature by explaining why and under what conditions ensembles improve algorithmic fairness in high-stakes applications.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 18133