Author(s) :

Zihao Ye.

Volume/Issue :

Abstract :

This study examines whether machine learning models trained on aggregated institutional data exhibit varying prediction accuracy among different school performance groups within Ontario’s public education system. Utilizing theoretical frameworks from algorithmic fairness research, the study investigates whether predictive reliability exhibits systematic variation across educational contexts, even in the absence of explicit demographic variables. Using publicly available standardized test data from the Education Quality and Accountability Office, the study created regression models to predict how well schools perform academically based on their achievement levels. Model performance was evaluated using error-based metrics and subgroup error analysis across performance strata. Results indicate that prediction accuracy was not uniform across school groups, with lower-performing schools consistently exhibiting higher prediction error across model configurations. These results show that predictive models created from combined educational data might perform differently depending on the specific context. This underscores the necessity of assessing predictive consistency across institutional contexts when implementing machine learning techniques in education.