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 math used to define fairness is often "jagged" and impossible for standard optimization algorithms to handle directly.
In their paper, "SLIDE: a surrogate fairness constraint to ensure fairness consistency," researchers from Seoul National University and Samsung Electronics propose a novel solution to this problem that is both mathematically rigorous and practically superior to existing methods.
TL;DR
The authors introduce SLIDE, a new surrogate function for fairness constraints that replaces non-differentiable indicator functions. Unlike standard convex surrogates (like the hinge loss), SLIDE is theoretically guaranteed to be "fairness-consistent," meaning it actually converges to the optimal fair model as more data is provided, resulting in a better trade-off between accuracy and fairness.
What Problem Does This Paper Address?
Most fairness definitions (like Disparate Impact) rely on the indicator function—a mathematical "switch" that is either 0 or 1. Because this function has no gradient (it’s flat everywhere except at zero, where it’s undefined), you can't use standard gradient descent to train a model with it.
To get around this, researchers typically use surrogate constraints. The most common is the hinge surrogate, which provides a smooth, convex upper bound to the indicator function. However, the authors point out a critical flaw: the hinge surrogate doesn't necessarily lead to the same optimal model as the original fairness constraint. In technical terms, it lacks fairness consistency. Using a hinge surrogate is like trying to hit a bullseye by aiming at a larger, slightly shifted target—you might get close, but you’ll likely miss the center.
Methods
The core contribution is the SLIDE function ($\nu_\tau$). Instead of being an upper bound like the hinge loss, SLIDE is a lower bound that looks like a playground slide: it starts at 0, slopes up linearly to 1 over a small interval ($\tau$), and then stays at 1.
Key Technical Features:
- Fairness Consistency: The authors prove that as the sample size $n$ grows, a model trained with SLIDE will satisfy the original fairness constraint asymptotically.
- Non-Convexity: Unlike the hinge loss, SLIDE is non-convex. While this sounds daunting, the authors show that modern optimizers like Adam handle it well. They also provide a CCCP (Concave-Convex Procedure) algorithm for those who want more stable convergence.
- Uniform Individual Fairness (UIF): The paper proposes a new definition for individual fairness. While traditional definitions look at the average difference in treatment between similar people, UIF looks at the probability that similar people are treated differently, which the authors argue is a more robust metric.
Datasets
The authors evaluated SLIDE against state-of-the-art baselines using several standard fairness benchmarks:
- Adult Dataset: Census data used to predict whether income exceeds $50k/yr (sensitive variable: gender).
- Bank Dataset: Marketing data used to predict if a client will subscribe to a term deposit (sensitive variable: age).
- Law Dataset: Law school admission data used to predict GPA (sensitive variable: race).
- Toy Datasets: 2D Gaussian mixtures and "two-moon" datasets were used to visualize how different constraints warp the decision boundary.
Key Insights and Ideas
- The "Lower Bound" Advantage: By using a lower bound (SLIDE) rather than an upper bound (Hinge), the optimization process is less "choked" by the constraint. This allows the model to find solutions that are both fairer and more accurate.
- Consistency Matters: The paper provides a rigorous theoretical bridge between surrogate optimization and population-level fairness. They derive convergence rates showing that SLIDE doesn't just work in practice—it is mathematically destined to reach the right answer.
- Pareto Dominance: In multi-objective optimization, one method "dominates" another if it is better in at least one metric without being worse in others. The authors show that SLIDE often dominates hinge-based methods, providing a better "Pareto front" of accuracy vs. fairness.
Key Results
- Better Fairness: On the Adult dataset, SLIDE achieved a Disparate Impact (DI) of 0.007, significantly lower than the Hinge surrogate's 0.013, while maintaining comparable accuracy.
- Individual Fairness: Using the new UIF definition, SLIDE outperformed specialized individual fairness algorithms like SenSR and SenSeI on consistency metrics (e.g., achieving a consistency score of 0.970 vs. SenSR's 0.934).
- Robustness to Hyperparameters: The performance of SLIDE was found to be relatively stable across different values of $\tau$ (the width of the "slide"), making it easier to tune in practice.
Limitations and Future Work
- Optimization Complexity: Because SLIDE is non-convex, it can be sensitive to initial conditions. The authors suggest a "Hybrid SLIDE" (HySLIDE) approach—starting with a hinge-trained model and then refining it with SLIDE—to mitigate this.
- Computational Cost: Calculating individual fairness (UIF) requires generating adversarial examples during training, which increases the computational burden compared to simpler group fairness metrics.
- Future Directions: The authors suggest that the SLIDE approach could be extended to other complex fairness definitions beyond DI and UIF, and potentially used to create better fairness evaluation tools, not just training constraints.
Summary updated