OT-FairBoost: Optimal Transport-Guided Gradient Boosting for Fairness Regularization on Tabular Data

Tree-based models like gradient boosting remain widely used for real-world applications despite recent attention to neural networks, meaning fairness solutions for these models are practically important for addressing algorithmic bias in deployed AI systems.

OT-FairBoost introduces an "in-processing" fairness method that works during model training rather than before or after, by adding a fairness penalty directly into gradient boosting's objective function—this approach is less common for tree-based models compared to neural networks.

The method targets specific fairness criteria including Demographic Parity (equal prediction rates across groups) and Equalized Odds (equal error rates across groups), providing organizations with measurable fairness standards to work toward.

The approach was tested across multiple realistic scenarios including binary classification, regression tasks, and situations with multiple demographic groups, suggesting it can be applied to diverse real-world prediction problems beyond simple two-group comparisons.

This work addresses a gap in fairness tools available for gradient boosting, which is important since many organizations use these models in high-stakes decisions (lending, hiring, etc.) where fairness and regulatory compliance are critical concerns.