Abstract
As artificial intelligence becomes increasingly prevalent in education, ensuring educational fairness has emerged as a critical concern. Algorithmic bias can lead to inequitable predictions, resulting in the unfair allocation of educational resources and posing significant social and ethical challenges. Therefore, designing and evaluating fair educational algorithms is a pressing challenge. This paper introduces FairEduNet, a novel framework for fairness optimization that combines a Mixture of Experts (MoE) architecture with adversarial training. The MoE architecture enhances prediction accuracy, while the adversarial network systematically reduces the model's dependence on sensitive attributes. This dual-component design allows FairEduNet to significantly mitigate implicit bias while maintaining high predictive accuracy for student dropout prediction. We evaluate FairEduNet on three educational datasets-the Duolingo dataset, the Portuguese Student Performance dataset, and the Open University Learning Analytics Dataset-comparing its performance against four debiasing algorithms. Experimental results demonstrate that FairEduNet achieves a significant improvement in fairness metrics across these datasets without compromising predictive accuracy.