Mitigating Bias in Tree-Based Models: A Fairness Perspective

This talk presents the Fairness Uncertainty-Aware Classification Tree (FUNACT), a novel in-processing framework that extends the classical CART algorithm to jointly optimize predictive accuracy and algorithmic fairness. The proposed approach explicitly accounts for the statistical uncertainty associated with fairness estimation by constructing confidence intervals for group fairness metrics, such as statistical parity, at each candidate split. Splits exhibiting statistically significant fairness violations, identified when the corresponding confidence interval excludes zero, are penalized through a graduated adjustment of the Information Gain criterion. A tunable regularization parameter allows us to flexibly balance predictive performance and fairness according to the application context. Experimental results on both synthetic and benchmark datasets show that FUNACT effectively mitigates algorithmic bias while preserving the transparency, interpretability, and computational efficiency of decision trees.