Feature Selection Analysis for Student Dropout Early Warning System Using ADASYN and XGBoost Methods

Student dropout remains a critical challenge in higher education, particularly when class distributions are highly imbalanced. This research proposes a hybrid framework combining ADASYN oversampling with XGBoost classification, validated using a split-first resampling protocol to prevent data leakage into the test set. The framework is evaluated on two datasets: a local Indonesian university dataset (UMBJM, n = 4,923) and the UCI “Predict Students’ Dropout and Academic Success” dataset (n = 4,424). A combined resampling strategy, majority undersampling paired with ADASYN minority oversampling, is applied exclusively to training folds. Results demonstrate that the proposed method achieves 94% recall on the UMBJM dataset, compared to a 15% baseline, and 91% recall on the UCI dataset, compared to a 62% baseline. These improvements are statistically significant (p-values < 0.01) based on paired t-tests across 5-fold cross-validation; the 95% confidence intervals for recall gain are 0.38–0.52 (UMBJM) and 0.25–0.33 (UCI), indicating substantial practical significance. Feature importance analysis reveals context-specific predictors: academic performance dominates in the local dataset, whereas socioeconomic and macroeconomic indicators are more influential in the international dataset. These findings support developing institution-specific early warning systems.