Evaluating Algorithmic Bias in Models for Predicting Academic Performance of Filipino Students

Algorithmic bias in education is understudied in Asian contexts: This research addresses a significant gap by being among the first to systematically examine how machine learning models for predicting student performance may unfairly treat students differently based on regional background in an Asian educational setting.

The study examined a substantial real-world dataset: Researchers analyzed nearly 49 million learning activity records from 5,986 Filipino students across three semesters, providing robust evidence based on actual student behavior rather than simulated data.

No regional bias was detected in this particular model: Using three different fairness metrics (AUC, weighted F1-score, and MADD), the researchers found that the grade prediction model performed consistently across all regional groups, suggesting the model did not systematically disadvantage students from any particular region.

Fairness evaluation requires multiple metrics: The paper demonstrates that responsible AI assessment should use several complementary fairness measures rather than relying on a single metric, as different metrics can reveal different types of bias.

Context-specific bias research is essential: While this study found no regional bias in one Philippine university's system, the findings highlight the importance of testing for algorithmic bias in specific educational contexts rather than assuming results from other regions or countries will apply universally.