Trade-offs between fairness and performance in educational AI: Analyzing post-processing bias mitigation on the OULAD

Abstract

Context:

AI-driven educational tools often face a trade-off between fairness and performance, particularly when addressing biases across sensitive demographic attributes. While fairness metrics have been developed to monitor and mitigate bias, optimizing all of these metrics simultaneously is mathematically infeasible, and adjustments to fairness often result in a decrease in overall system performance.

Objective:

This study investigates the trade-off between predictive performance and fairness in educational AI systems, focusing on gender and disability as sensitive attributes. We evaluate whether post-processing fairness interventions can mitigate group-level disparities while preserving model usability.

Method:

Using the Open University Learning Analytics Dataset, we trained four machine learning models to predict student outcomes. We applied the equalized odds post-processing technique to mitigate bias and assessed model performance with accuracy, F1-score, and AUC, alongside fairness metrics including statistical parity difference (SPD) and equal opportunity difference (EOD). Statistical significance of changes was tested using the Wilcoxon signed-rank test.

Results:

All models achieved strong baseline predictive performance, with RF performing best overall. However, systematic disparities were evident, particularly for students with disabilities, showing that high accuracy does not necessarily ensure equitable outcomes. Post-processing reduced group-level disparities substantially, with SPD and EOD values approaching zero, though accuracy and F1-scores decreased slightly but significantly. RF and ANN were more resilient to fairness adjustments.

Conclusion:

This study highlights the importance of fairness-aware machine learning, such as post-processing interventions, and suggests that appropriate mitigation methods should be used to ensure benefits are distributed equitably across diverse learners, without favoring any particular fairness metric.