Gradient descent method learns fairness metrics from simple comparisons

Individual fairness, the principle that 'similar individuals should be treated similarly,' offers a strong guarantee for algorithmic decision-making but faces a major practical barrier: learning the similarity metric over individuals. A new paper from Conlan Olson and colleagues tackles this head-on with an algorithm that learns a Mahalanobis distance metric from simple triplet comparisons. Users provide queries like 'is person A more similar to person B or C?' and the system fits a Bradley-Terry pairwise comparison model. The algorithm starts with a spectral initialization to get a good starting point, then applies gradient descent to refine the metric. The authors provide extensive theoretical guarantees, proving the non-convex optimization converges quickly to the ground truth metric. They also show that ensuring fairness under the learned metric is sufficient to achieve near-identical fairness under the true metric, closing the loop between estimation and application.

The work has direct implications for deploying fair AI systems in practice. By reducing the burden of specifying similarity metrics manually, it allows models to be fine-tuned for fairness using only human or automated triplet judgments. The authors demonstrate convergence and downstream fairness performance through experiments. This approach could be integrated into model tuning pipelines, making individual fairness operational for industries like hiring, lending, or healthcare. The paper (arXiv:2605.23145, 60 pages) is a significant step toward turning a theoretical ideal into a deployable tool for AI ethics.