- Alyssa Whitlock Lees
- Ananth Balashankar
Prevalent methodology towards constructing fair machine
learning (ML) systems, is to enforce a strict equality metric for de- mographic groups based on protected attributes like race and gender.
While definitions of fairness in philosophy are varied, mitigating bias in ML classifiers often relies on demographic parity-based constraints across sub-populations. However, enforcing such constraints blindly can
lead to undesirable trade-offs between group-level accuracy if groups pos- sess different underlying sampled population metrics, an occurrence that
is surprisingly common in real-world applications like credit risk and income classification. Similarly, attempts to relax hard constraints may lead to unintentional degradation in classification performance, without
benefit to any demographic group. In these increasingly likely scenar- ios, we make the case for transparent human intervention in making the
trade-offs between the accuracies of demographic groups. We propose that transparency in trade-offs between demographic groups should be
a key tenet of ML design and implementation. Our evaluation demon- strates that a transparent human-in-the-loop trade-off technique based
on the Pareto principle increases both overall and group-level accuracy by 9.5% and 9.6% respectively, in two commonly explored UCI datasets for credit risk and income classification.