Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2026
- Goals
- Record
- ef1126cd-c07e-4cc9-b971-959aa827672e
- Published
- 2026-10-30
Universitas Dian Nuswantoro (UDINUS) conducted a study presenting Hybrid Logistic Regression Stacking (HLR-STACK), a hybrid stacking framework for predicting student outcomes in higher education. The framework uses logistic regression as the meta-learner and integrates leakage-safe preprocessing, infold Synthetic Minority Over-sampling Technique (SMOTE) for class balancing, and out-of-fold stacking over tree-based and boosting base learners.
The activity considered two prediction tasks. The first was binary classification distinguishing Graduate and Dropout outcomes. The second was multiclass classification involving Dropout, Enrolled, and Graduate outcomes. Top-K ensemble selection was used to maintain a compact and deployment-friendly model. Evaluation on a 4,424-record dataset showed that HLR-STACK achieved 0.981 accuracy for the binary task and 0.844 accuracy for the multiclass task, outperforming single-model baselines.
The study also addressed interpretable probability-based risk scoring. Post-hoc calibration using Maximum Calibration Error (MCE) and reliability diagrams supported interpretation of the model outputs. Recursive Feature Elimination (RFE) identified the top 30 predictors, highlighting semester pass rates, study load, grades, and course structure as primary contributors.
Through this activity, UDINUS presented a practical, auditable, scalable, and interpretable decision-support workflow for early identification of at-risk students. The source describes the workflow as facilitating targeted interventions and evidence-based educational policies. HLR-STACK integrates complementary base learners through a transparent meta-learner and produces interpretable outputs. The framework can be extended to other higher education institutions and datasets, providing a basis for student outcome prediction, retention-related analysis, and academic performance prediction.