Adaptive Pareto Exploration (APEX) for Fairness-Aware Hyperparameter Optimization in FairPilot

Context:

The integration of machine learning (ML) into high-stakes software systems, such as those used in education, criminal justice, and finance, has elevated concerns over fairness, transparency, and accountability. Traditional hyperparameter optimization approaches often overlook fairness considerations, limiting their suitability for responsible AI development.

Objectives:

This work introduces APEX, a fairness-aware MOBO algorithm, and FairPilot, an interactive system for exploring accuracy–fairness trade-offs. We aim to enable practitioners to explore and operationalize fairness-aware ML configurations more effectively.

Methods:

APEX, integrates a Pareto-based coordinate selection strategy and a perturbation mechanism that prioritizes hyperparameters based on their joint influence on fairness and accuracy. Built on that, FairPilot visualizes trade-offs between fairness and accuracy across ML models and metrics. We evaluate the system across diverse datasets (e.g., COMPAS, ELS, German Credit) and multiple model types.

Results:

APEX consistently outperforms baseline optimization strategies such as ParEGO and EHVI, achieving higher hypervolume coverage and converging more quickly to fairness–accuracy trade-offs of superior quality. An analysis of hyperparameter importance reveals that regularization parameters play a central role in improving fairness while preserving predictive performance.

Conclusions:

FairPilot and APEX jointly provide a novel, user-centered, and algorithmically grounded approach to fairness-aware ML. By supporting both visual decision-making and targeted optimization, our system facilitates responsible model development, offering a practical and extensible solution for fairness in software systems.