Abstract
As they have a vital effect on social decision makings, AI algorithms should be not only accurate and but also fair. Among various algorithms for fairness AI, learning a prediction model by minimizing the empirical risk (e.g., cross-entropy) subject to a given fairness constraint has received much attention. To avoid computational difficulty, however, a given fairness constraint is replaced by a surrogate fairness constraint as the 0-1 loss is replaced by a convex surrogate loss for classification problems. In this paper, we investigate the validity of existing surrogate fairness constraints and propose a new surrogate fairness constraint called SLIDE, which is computationally feasible and asymptotically valid in the sense that the learned model satisfies the fairness constraint asymptotically and achieves a fast convergence rate. Numerical experiments confirm that the SLIDE works well for various benchmark datasets.
Introduction
AI models are increasingly used for high-stakes decisions like credit scoring and criminal risk assessment. However, these models can inherit historical biases, leading to unfair treatment of sensitive groups. While researchers have developed "in-processing" methods to bake fairness directly into the training process, a major technical hurdle remains: the mathematical definitions of fairness often rely on non-differentiable indicator functions—mathematical "switches" that are either 0 or 1. Because these functions have no gradient, standard optimization algorithms like gradient descent cannot handle them directly.
The SLIDE Surrogate Function
To circumvent this, researchers typically use surrogate constraints, most commonly the hinge surrogate. However, the hinge surrogate provides an upper bound that often lacks "fairness consistency," meaning it may not converge to the optimal fair model as more data is provided.
The authors propose SLIDE ($\nu_\tau$), a novel surrogate function that acts as a lower bound to the indicator function. As illustrated in Figure 4, SLIDE starts at 0, slopes up linearly to 1 over a small interval ($\tau$), and then remains at 1. This "slide" shape allows for smooth optimization while maintaining theoretical guarantees of fairness consistency. Although SLIDE is non-convex, the authors demonstrate that modern optimizers like Adam, or specialized algorithms like the Concave-Convex Procedure (CCCP), can effectively find high-quality solutions.
Experimental Evaluation
The authors evaluated SLIDE against state-of-the-art baselines using standard fairness benchmarks, including the Adult, Bank, and Law datasets. They introduced a new metric, Uniform Individual Fairness (UIF), which measures the probability that similar individuals are treated differently, providing a more robust measure than traditional average-based metrics.
The results, summarized in Figure 6, show that SLIDE often "Pareto dominates" hinge-based methods, providing a superior trade-off between accuracy and fairness. On the Adult dataset, for example, SLIDE achieved a Disparate Impact (DI) of 0.007, significantly lower than the Hinge surrogate's 0.013, while maintaining comparable accuracy. SLIDE also outperformed specialized individual fairness algorithms like SenSR and SenSeI on consistency metrics.
Conclusion and Future Work
SLIDE provides a mathematically rigorous bridge between surrogate optimization and population-level fairness. While its non-convexity can introduce sensitivity to initial conditions—a challenge the authors address with a "Hybrid SLIDE" (HySLIDE) approach—its theoretical consistency and empirical performance make it a powerful tool for fair AI. Future work may extend this approach to more complex fairness definitions and evaluation tools.
Summary updated