Authors - Caio Tertuliano Ribeiro, Lilian Berton Abstract - This paper studies a Disparate Impact (DI)-oriented variant of fairness-aware hyperparameter optimization for XGBoost in banking and credit settings. The method combines Design of Experiments (DoE), Response Surface Methodology (RSM), and NBI-style sampling to jointly tune XGBoost hyperparameters, the decision threshold, and the positive-class weight under a fixed evaluation budget. Unlike the preliminary composite-fairness draft, the final experiment optimizes a DI-only objective while keeping Statistical Parity Difference (SPD), Equal Opportunity Difference (EOD), and Average Odds Difference (AOD) as audit metrics. Experiments on Bank Marketing, German Credit, and Default of Credit Card Clients with 30 replicas per dataset show that the proposed utopia selector is competitive with evaluation-matched random search and consistently superior to the XGBoost default configuration. Relative to the default baseline, it improves Balanced Accuracy by +0.059, +0.020, and +0.019, while reducing the DI gap by -0.375, -0.174, and -0.141, respectively. The main takeaway is that DI-only optimization provides a finance-oriented and reproducible way to navigate fairness–performance trade-offs rather than a universal domination claim over random search.
Open Zoom