STUDENT PERFORMANCE PREDICTION USING MACHINE LEARNING

The problem of early prediction of at-risk students is an essential issue in contemporary education frameworks. In this research paper, we propose a comprehensive machine learning pipeline for predicting student academic performance on the OULAD dataset containing information about 32,593 learners enrolled in 22 different classes. In this work, we compare four models – XGBoost, LightGBM, Random Forest, and Logistic Regression – assess the influence of domain-oriented features engineering and Synthetic Minority Over-sampling Technique (SMOTE) data balance strategy, and employ Optuna for Bayesian hyperparameter optimization of the selected models. The XGBoost model with Optuna optimization, shows 97% test AUC and 91% test F1 score. An ablation experiment estimates the significance of each step of our pipeline, while SHapley Additive exPlanations (SHAP) values deliver interpretable results and explain individual predictions in the context understandable by educators.