Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2026
- Goals
- Record
- ff1af430-8826-46be-b241-bfacd9923343
- Published
- 2026-10-30
Universitas Dian Nuswantoro conducted research on predicting student academic performance in Learning Management Systems (LMS), involving researchers Hermaliani, Eni Heni; Fanani, Ahmad Zainul; Santoso, Heru Agus; and Affandy as activity participants. The activity addressed challenges associated with class imbalance and the black-box nature of many high-accuracy prediction models, which can limit their practical adoption in educational settings.
The activity proposed an explainable stacking ensemble framework that combined heterogeneous base learners, namely k-nearest neighbors, random forest, extra trees, XGBoost, and LightGBM, with Logistic Regression as a meta-learner for transparent aggregation. A hybrid SMOTE-ENN resampling strategy was applied exclusively within training folds under stratified five-fold cross-validation to address class imbalance and prevent data leakage. Model interpretability was supported through Shapley Additive Explanations, providing global feature importance and local prediction explanations.
Experiments used the Kalboard 360 benchmark dataset, comprising 480 instances and 16 features, with three performance levels: High, Medium, and Low. The framework achieved an accuracy of 96.57%, precision of 96.34%, recall of 93.26%, and an F1-score of 94.40%, outperforming individual classifiers and homogeneous ensembles. Low fold-wise variance confirmed stability across validation splits. Wilcoxon signed-rank tests at an alpha level of 0.05 found no statistically significant difference between the proposed model and baseline classifiers, although the stacking framework consistently achieved the highest mean scores.
SHAP analysis showed that behavioral engagement features, particularly student absence days and visited resources, dominated predictions, while demographic attributes contributed negligibly. The activity concerns student performance prediction, early warning analytics, and educator-oriented intervention strategies in LMS-based educational information systems. The research used leakage-safe resampling and SHAP-based explainability to provide a practical and interpretable solution for student performance prediction.