Bias-Conscious Splitting: A Novel Classification Tree for Fair Decision-Making

We introduce an uncertainty-aware version of the CART algorithm designed to promote fairer predictions while preserving strong predictive performance. Our approach leverages an in-processing technique that directly integrates fairness considerations into the model’s learning process. Specifically, 1) it uses Disparate Impact as a common measure to assess the fairness of decision-making models, and 2) it adjusts the loss function of the classification tree to account for uncertainty in the fairness metric. More precisely, the algorithm evaluates each potential split by constructing confidence intervals for the fairness metric—when these intervals exclude one, indicating significant discrimination, our method applies a graduated penalty that scales with the degree of unfairness. This penalty reduces the Information Gain more severely for highly discriminatory splits, naturally steering the algorithm toward more equitable decision boundaries.