WattsRight: An Empirical Investigation of Data-Centric Interventions in Algorithmic Fairness Assessment

While algorithmic bias awareness has increased,
fairness assessment typically occurs post-deployment rather than
being embedded throughout the development life-cycle. Current
tools for fairness assessment face adoption challenges due to steep
learning curves and complex interpretation requirements. Beyond
these technical barriers, decision-makers frequently lack clear
frameworks for how to approach fairness implementation and
what choices are available to them throughout the development
process. To address this need, we propose an interactive dash-
board that examines how feature selection decisions, resource
constraints, and correlations between protected attributes and
other features influence fairness outcomes across demographic
groups. This visualization tool focuses on three group-fairness
metrics: Demographic Parity, Equalized Odds, and Predictive
Parity, while also illustrating the inherent trade-offs between
these metrics. The dashboard enables users to experiment with
data-centric interventions including feature selection and bud-
geting to understand their impact on fairness. The goal is to
create a clear visualization tool that raises awareness for stake-
holders responsible for AI system development and oversight,
including data scientists, product managers, and compliance
officers, allowing them to see how certain choices might benefit or
negatively affect demographic groups. Evaluation demonstrates
these interventions significantly affect fairness outcomes, with
stakeholders rating the tool positively for practical utility.