A comparative study of manual and machine learning approaches for predicting student academic performance using extreme gradient boosting with shapley additive explanations

Abstract

Predicting student performance is vital in order to make appropriate decisions and intervene in time. The main goal of the present study is to develop an explainable model using Shapley Additive Explanations and Extreme Gradient Boosting techniques, and to evaluate its performance against a traditional manual rule-based model. The problem is considered a binary classification problem. The data was split into training and testing subsets. The manual rule-based model achieves an F1 score of 90.10%, a recall of 82.73%, an accuracy of 84.62%, and precision of 98.91%. The Extreme Gradient Boosting model achieved higher point estimates than the manual rule-based model, with 90.77% accuracy, 93.75% precision, 95.45% recall, and 94.59% F1-score, although McNemar's test indicated that this difference was not statistically significant (p = 0.115), and the manual model identified at-risk (Fail-class) students more reliably (95.00% recall) than XGBoost (65.00% recall). The Shapley Additive Explanations analysis shows that the grades obtained in the first and second grading periods are the most important factors in determining academic performance.