Artificial Intelligence for Academic Decision-Making: Predicting Student Performance Using Ensemble Learning
TL;DR
An ensemble-learning framework that could predict the final course grades from the UCI Student Performance dataset in a leakage-resistant manner is created, illustrating how ensemble learning can be used to provide interpretable risk-oriented academic decision support and maintain clear boundaries for causal interpretations and generalizability from student records.
Abstract
AI can assist in academic decision making to help pinpoint students who might need extra attention in their education at an opportune time. This study created an ensemble-learning framework that could predict the final course grades from the UCI Student Performance dataset in a leakage-resistant manner. There were no missing or duplicate records in the Portuguese-language cohort (649 students and 33 variables). Two information scenarios were assessed: an early-warning scenario (no interim grades G1 and G2) and an updated-performance scenario (with interim grades G1 and G2). The four algorithms (Ridge regression, Random Forest, Extra Trees and Gradient Boosting) were tuned on the development data and evaluated on an independent validation partition. Extra Trees was evaluated once on the locked test set since it had the best validation performance (RMSE - 1.290, MAE - 0.776, R2 - 0.864) among the models trained in the updated-performance scenario. It obtained RMSE 1.254, MAE 0.804, and R2 0.811. The permutation analysis result indicated that G2 was the strongest predictor, and the risk-band analysis result indicated that the band accuracy was 88.89% for the high-risk students and 89.47% for the high-performing students. The framework illustrates how ensemble learning can be used to provide interpretable risk-oriented academic decision support and maintain clear boundaries for causal interpretations and generalizability from student records.