EXPLAINABLE MACHINE LEARNING FOR STUDENT ACADEMIC PERFORMANCE PREDICTION IN DATA-CONSTRAINED EDUCATIONAL SETTINGS

Abstract

The research explores the effect of a limited dataset in terms of size on machine learning models used for predicting student academic performance in schools. Model generalization can easily be achieved when the dataset size is large. This is not the case in educational data mining, due to the limited data in the field, making model generalization, interpretability, and fairness difficult to achieve. A stratified subsampling technique was used at sizes of 50, 100, 150, 200, and 300 to handle the data scarcity problem. In the same vein, an oversampling strategy was used to balance classes that were not well represented to eliminate bias. SHAP and permutation importance were used to perform interpretability of results, while Spearman rank correlation and Jaccard similarity were used for explanation stability. A fairness audit was also carried out to identify how other socio-demographic factors, other than gender, affect academic performance. The dataset used in this research is the mathematics scores of Portuguese students. Standardized scaling, ordinal encoding, one-hot encoding, and target attribute definition (pass or fail) are performed on the dataset. Logistic Regression, Random Forest, and XGBoost models were trained and evaluated on the dataset. The results from the models were evaluated using accuracy, F1-score, AUC, and Brier score metrics. Results show that Random Forest and XGBoost performed better in terms of accuracy, robustness, and calibration, even with small datasets, when compared to the Logistic Regression model.