Towards Equity and Algorithmic Fairness in Student Grade Prediction

Algorithmic bias in grade prediction is measurable and addressable: The researchers found that standard grade prediction algorithms perform differently across racial groups, but they successfully tested multiple strategies to reduce these disparities—demonstrating that fairness in educational AI is not impossible, just requires intentional design choices.

Adversarial learning combined with balanced training data works best: The study found that an "adversarial learning approach" paired with balanced grade labels was most effective at achieving fair predictions across racial groups, suggesting that institutions should consider this technical combination when deploying grade prediction systems.

Equity requires intentional oversampling of underserved groups: To improve predictive accuracy specifically for historically underserved student populations, the researchers found success by deliberately including more data from these groups during model training—a concrete strategy institutions can implement to prevent AI systems from perpetuating existing educational inequities.

Performance trade-offs must be transparent: The paper emphasizes that institutions need frameworks to understand and openly discuss the trade-offs between overall prediction accuracy and fairness across racial groups, rather than optimizing for one metric without considering impacts on equity.

AI grade prediction tools need equity guardrails before campus deployment: As colleges increasingly use AI for academic support decisions, this research provides methodologies institutions should adopt to ensure these systems don't inadvertently disadvantage students based on race.