A Human-in-the-Loop Fairness-Aware Model Selection Framework for Complex Fairness Objective Landscapes

TL;DR

This work introduces ManyFairHPO, a human-in-the-loop, fairness-aware model selection framework that enables practitioners to effectively navigate complex and nuanced fairness objective landscapes and aids in the identification, evaluation, and balancing of fairness metric conflicts and their related social consequences, leading to more informed and socially responsible model-selection decisions.

Abstract

Fairness-aware Machine Learning (FairML) applications are often characterized by complex social objectives and legal requirements, frequently involving multiple, potentially conflicting notions of fairness. Despite the well-known Impossibility Theorem of Fairness and extensive theoretical research on the statistical and socio-technical trade-offs between fairness metrics, many FairML tools still optimize or constrain for a single fairness objective. However, this one-sided optimization can inadvertently lead to violations of other relevant notions of fairness. In this socio-technical and empirical study, we frame fairness as a many-objective (MaO) problem by treating fairness metrics as conflicting objectives. We introduce ManyFairHPO, a human-in-the-loop, fairness-aware model selection framework that enables practitioners to effectively navigate complex and nuanced fairness objective landscapes. ManyFairHPO aids in the identification, evaluation, and balancing of fairness metric conflicts and their related social consequences, leading to more informed and socially responsible model-selection decisions. Through a comprehensive empirical evaluation and a case study on the Law School Admissions problem, we demonstrate the effectiveness of ManyFairHPO in balancing multiple fairness objectives, mitigating risks such as self-fulfilling prophecies, and providing interpretable insights to guide stakeholders in making fairness-aware modeling decisions.

Figures (7)

• Figure 1: ManyFairHPO: We introduce ManyFairHPO, a human-in-the-loop, many-objective optimization framework for navigating complex fairness and performance trade-offs in machine learning. The approach consists of three main stages: 1) Many-Objective Optimization to explore trade-offs between a candidate set of fairness and performance metrics, 2) Fairness Metric Selection & Risk Identification to incorporate domain knowledge, human preferences, and insights from Pareto front analysis to prioritize metrics and assess fairness conflict risks, and 3) Many-Objective Model Selection to choose a final model that balances multiple objectives based on assigned weights.
• Figure 2: Conflict-Related Risks: Fairness metrics serve as a statistical means to specify and optimize various social objectives, such as equality, equity, and individual justice. However, certain conflicts hold potential downstream consequences, such as the risk of Self-Fulfilling Prophecy (SFP) when DDSP is satisfied by violating other metrics.
• Figure 3: ManyFairHPO vs. Bi-Objective Optimization: Comparison of hypervolume achieved by ManyFairHPO and bi-objective optimization across datasets and search spaces. NSGA-III matches the hypervolume achieved by NSGA-II with a correlation of 0.991, indicating that similar fairness-accuracy Pareto Fronts are achieved when optimizing for fairness metrics together compared to optimizing for them separately.
• Figure 4: Fairness Metric Contrast: Heatmap visualization of fairness metric conflicts across different datasets and ML models, measured by the degree to which optimizing for one fairness-accuracy trade-off fails to optimize for another. Light red cells indicate more severe conflicts. The results highlight the data and model dependence of fairness metric conflicts.
• Figure 5: Conflict Interpolation (RF-Adult): Ternary plot of the high-dimensional Pareto Front (orange) in the presence of a fairness metric conflict between INVD (purple) and DEOP (green). Metric values get better further away from each corner. The many-objective Pareto Front provides a selection of models that interpolate between fairness metric conflicts.

Theorems & Definitions (2)

• Definition 3.1: FairML Problem
• Definition 4.1: Contrast