Introduction
As machine learning models are increasingly deployed in high-stakes domains such as criminal recidivism prediction and hiring, the risk of inheriting and amplifying historical human biases has become a critical concern. While various fairness constraints have been proposed, they often lead to non-convex optimization problems that are computationally challenging and lack clear probabilistic interpretations. This paper, presented at AAAI-18, introduces a novel "weighted sum of logs" technique that ensures convexity while providing a rigorous framework for training non-discriminatory classifiers.
The Weighted Sum of Logs Technique
The core innovation is a convex optimization approach designed to maximize the weighted probability of favorable outcomes for historically disadvantaged groups.
- Objective Function: Instead of standard error minimization, the authors maximize a weighted sum of the logarithms of probabilities for "favorable" outcomes (e.g., not recidivating or being hired).
- Dynamic Weighting: Weights ($w$) are assigned based on historical bias in the training data. Individuals from historically disfavored groups receive higher weights to counteract systemic bias.
- Loss Constraint: To maintain utility, the optimization is subject to a constraint ensuring the loss of the fair model does not exceed that of a standard "vanilla" model by more than a user-defined threshold ($\delta$).
- Convexity and Interpretability: By utilizing the logarithm of probabilities, the problem remains convex, allowing for efficient global optimization. Furthermore, this approach preserves the model's probabilistic integrity, facilitating human-understandable explanations for predictions.
Experimental Evaluation
The proposed method was evaluated on two benchmark datasets for fairness:
- ProPublica’s COMPAS Dataset: Predicting criminal recidivism with "Race" as the sensitive attribute.
- UCI Adult Income Dataset: Predicting high-income status with "Gender" as the sensitive attribute.
The primary metric for fairness is the $p$-rule, which measures the ratio of the probability of a favorable outcome for the unprivileged group versus the privileged group.
Results and Discussion
The empirical results demonstrate that the weighted sum of logs technique effectively navigates the inherent trade-off between accuracy and fairness. The authors prove (Proposition 3) that in a biased world, it is mathematically impossible to have a classifier that is both perfectly accurate and perfectly non-discriminatory; a trade-off is inevitable.
As shown in the figure above, the method achieves significant gains in fairness with minimal impact on predictive accuracy. In the COMPAS dataset, the $p$-rule improved from 64% to 90%, while accuracy only decreased from 67% to 66%. For the Adult Income dataset, the $p$-rule was boosted from 34% to 90%, with accuracy shifting from 85% to 82%.
The authors also introduce the concept of weighted proportional fairness, borrowing from communication network theory. A classifier is considered fair if no alternative can improve the relative favoring of one group without causing a disproportionately larger disadvantage to another.
Conclusion and Future Work
The "weighted sum of logs" technique provides a mathematically grounded and computationally efficient path toward non-discriminatory AI. By maintaining convexity and probabilistic integrity, it offers advantages over existing non-convex methods. While highly effective, the authors note that the choice of weights remains a policy decision. Future work includes extending this framework to more complex architectures like deep neural networks and addressing intersectional bias involving multiple sensitive attributes.
Summary updated