Beyond Statistical Fairness

In recent years, a great deal of fairness notions has been proposed. Yet, most of them take a reductionist approach by indirectly viewing fairness as equalizing some error statistic across pre-defined groups. This thesis aims to explore some ideas as to how to go beyond such statistical fairness frameworks. First, we consider settings in which the right notion of fairness may not be captured by simple mathematical definitions but might be more complex and nuanced and thus require elicitation from individual or collective stakeholders. By asking stakeholders to make pairwise comparisons to learn which pair of individuals should be treated similarly, we show how to approximately learn the most accurate classifier or converge to such one subject to the elicited fairness constraints. We consider an offline setting where the pairwise comparisons must be made prior to training a model and an online setting where one can continually provide fairness feedback to the deployed model in each round. We also report preliminary findings of a behavioral study of our framework using human-subject fairness constraints elicited on the COMPAS criminal recidivism dataset. Second, unlike most of the statistical fairness framework that promises fairness for pre-defined and often coarse groups, we provide fairness guarantees for finer subgroups, such as all possible intersections of the pre-defined groups, in the context of uncertainty estimation in both offline and online setting. Our framework gives uncertainty guarantees that are more locally sensible than the ones given by conformal prediction techniques; our uncertainty estimates are valid even when averaged over any subgroup, but uncertainty estimates in conformal predictions are usually only valid when averaged over the entire population.