Fairness-Aware Policy Gradient Methods for Mitigating Demographic Bias in AI-Driven Student Performance Prediction

Abstract

The proliferation of artificial intelligence systems in educational contexts has raised critical concerns regarding algorithmic fairness and demographic bias in student performance prediction. Traditional machine learning models, while achieving high predictive accuracy, often perpetuate or amplify existing disparities across gender, ethnicity, and socioeconomic backgrounds. This paper introduces a novel fairness-aware policy gradient framework that integrates demographic parity constraints into reinforcement learning architectures for student performance forecasting. Our approach employs constrained policy optimization to balance prediction accuracy with fairness metrics, utilizing a dual-objective reward function that penalizes discriminatory predictions while maintaining educational relevance. Through comprehensive experiments on diverse student datasets, we demonstrate that policy gradient methods with fairness constraints achieve superior bias mitigation compared to conventional supervised learning approaches, reducing demographic disparity by an average of 37.4% while preserving predictive performance within 2.8% of baseline models. The proposed framework incorporates adaptive fairness thresholds that dynamically adjust based on observed demographic distributions, enabling robust performance across heterogeneous educational environments. Our findings suggest that reinforcement learning-based fairness mechanisms provide a viable pathway for developing equitable AI systems in education while maintaining operational effectiveness for academic intervention strategies.