Dynamic Grade Prediction in Programming Education Using Time-Series XGBoost and SHAP Analysis

This study presents a time-series-based machine learning approach to predict final grades in a C programming course, leveraging temporal and behavioral features to support early intervention. We evaluated classification and regression models using three-class (Needs Improvement, Average, Excellent) and five-class (Fail, Poor, Average, Good, Excellent) grading schemes, addressing class imbalance with RandomOverSampler, SMOTE and ADASYN. Advanced sampling strategies, particularly SMOTE, enhanced minority class prediction in the three-class scheme, with XGBoost achieving superior performance. The five-class scheme offered finer granularity, revealing nuanced patterns in mid-tier performance through practice-related features, but faced challenges from increased class imbalance. Regression models, while suitable for continuous prediction, underperformed due to thresholding biases. SHAP analysis identified historical average score and difficulty-adjusted score as key predictors, providing actionable insights for educators. These findings highlight the trade-offs between broad and fine-grained prediction, with the three-class scheme supporting robust interventions and the five-class scheme enabling nuanced feedback. Future work includes incorporating qualitative features and hybrid approaches to improve fine-grained prediction and generalizability across educational contexts.