UNCOVERING HIDDEN BIASES IN MACHINE LEARNING MODELS: A STEP TOWARD ETHICAL AI

Abstract

With the widespread use of artificial intelligence (AI) systems in everyday applications, ensuring fairness has become a critical aspect of system design. AI models increasingly influence high-stakes decisions in domains such as finance, hiring, and criminal justice, where biased outcomes can have severe societal consequences. Recent developments in machine learning and deep learning have begun to address these challenges, emphasizing the need for systematic bias mitigation strategies.This study presents a structured framework for bias detection and mitigation in machine learning, aiming to reconcile predictive performance with algorithmic fairness. Using the UCI Adult dataset as a case study, the framework adopts a dual-phase approach: preprocessing with Reweighing to correct historical imbalances and postprocessing with Threshold Optimization to adjust decision thresholds for protected and unprotected groups. This combination addresses both data-level and decision-level disparities without requiring complete model retraining.The framework was evaluated across four supervised models—k-Nearest Neighbors, Decision Tree, Logistic Regression, and Random Forest. Mitigated models achieved substantial reductions in Demographic Parity Difference and Equal Opportunity Difference, alongside improved Disparate Impact Ratios, while maintaining competitive accuracy, precision, recall, and F1-score. These findings demonstrate that fairness-aware interventions can reduce group-level bias without critical performance loss, contributing to the development of trustworthy and socially responsible AI systems.