Gender Bias Detection and Mitigation in Income Prediction Using Machine Learning

Abstract

Algorithmic bias in machine learning systems is a major concern when predictive models are used in highstakes settings, including income prediction, hiring, healthcare, and resource allocation. This paper presents an empirical study of gender bias in six classification models Logistic Regression, Random Forest, Decision Tree, Support Vector Machine (SVM), Naive Bayes and XGBoost using the Census Income dataset. To complement standard group-level fairness metrics, a counterfactual sensitivity analysis is introduced in which the sensitive attribute sex is flipped at test time and the resulting prediction changes are summarized using the Counterfactual Flip Rate. It is emphasized that this procedure is a heuristic sensitivity test rather than a fully causal notion of counterfactual fairness. Three mitigation strategies are compared: reweighting, adversarial debiasing, and fairness constraints.