A Flexible Fairness Framework with Surrogate Loss Reweighting for Addressing Sociodemographic Disparities

TL;DR

The paper addresses the challenge of achieving sociodemographic fairness in predictive decision-making while preserving accuracy. It introduces α-β Fair Machine Learning (FML), a model-agnostic in-processing framework that embeds fairness directly into the learning objective via β-fair surrogate losses and group-wise loss reweighting, captured by the objective L_{(oldsymbol{α},oldsymbol{β})}(oldsymbol{w}). A distributed optimizer, Parallel Stochastic Gradient Descent with Surrogate Loss (P-SGD-S), is proposed along with convergence guarantees for both convex and nonconvex losses, enabling scalable training. Empirical results on datasets including Adult, COMPAS, and Fashion-MNIST demonstrate improved fairness-relevant metrics with competitive or superior average accuracy, and show that tuning β allows smooth interpolation between ERM and minimax fairness. The framework offers a flexible, principled approach to fairness that can adapt to diverse definitions (EA, DP, EO) and application contexts, with potential extensions to intersectional fairness and deeper models.

Abstract

This paper presents a new algorithmic fairness framework called $\boldsymbolα$-$\boldsymbolβ$ Fair Machine Learning ($\boldsymbolα$-$\boldsymbolβ$ FML), designed to optimize fairness levels across sociodemographic attributes. Our framework employs a new family of surrogate loss functions, paired with loss reweighting techniques, allowing precise control over fairness-accuracy trade-offs through tunable hyperparameters $\boldsymbolα$ and $\boldsymbolβ$. To efficiently solve the learning objective, we propose Parallel Stochastic Gradient Descent with Surrogate Loss (P-SGD-S) and establish convergence guarantees for both convex and nonconvex loss functions. Experimental results demonstrate that our framework improves overall accuracy while reducing fairness violations, offering a smooth trade-off between standard empirical risk minimization and strict minimax fairness. Results across multiple datasets confirm its adaptability, ensuring fairness improvements without excessive performance degradation.

Figures (10)

• Figure 1: Comparison of accuracy, loss, worst accuracy, and $\epsilon_{\text{EA}}$ for convex loss on Adult. Parallel SGD operates without explicit fairness constraints, while Minimax enforces an extreme fairness constraint by prioritizing the worst-performing group in group fairness.
• Figure 2: Comparison of accuracy, loss, worst accuracy, and $\epsilon_{\text{EA}}$ for nonconvex loss on Adult. Parallel SGD operates without explicit fairness constraints, while Minimax enforces an extreme fairness constraint by prioritizing the worst-performing group in group fairness.
• Figure 3: Impact of varying $\boldsymbol{\beta}$ values on training dynamics for convex loss on Adult. With $\beta_{1} = 0$, increasing $\beta_{0}$ prioritizes the minority group, improving worst-group accuracy and reducing $\epsilon_{\text{EA}}$.
• Figure 4: Comparison of accuracy, loss, worst accuracy, and $\epsilon_{\text{EA}}$ for convex loss on COMPAS. Parallel SGD operates without explicit fairness constraints, while Minimax enforces an extreme fairness constraint by prioritizing the worst-performing group in group fairness.
• Figure 5: Impact of varying $\boldsymbol{\beta}$ values on training dynamics for convex loss on COMPAS. With $\beta_{1} = 0$, increasing $\beta_{0}$ prioritizes the minority group, improving worst-group accuracy and reducing $\epsilon_{\text{EA}}$.

Theorems & Definitions (23)

• Definition 1: Equality of Accuracy (EA)
• Definition 2: EA violation
• Definition 3: $\beta$-fair Surrogate Loss Function
• Remark 1
• Definition 4: $\boldsymbol{\alpha}$-$\boldsymbol{\beta}$ Fair Surrogate Loss Function
• Remark 2
• Remark 3
• Remark 4
• Remark 5
• Remark 6