Explainable Ensemble Learning for Student Dropout Prediction in Conflict-Affected Educational Systems

This study investigates higher-education nonenrollment among secondary school graduates in Libya’s Fezzan region, formulated as a binary classification task distinguishing Continued Enrollment from Dropout Risk. A synthetic target variable was created using a rule-based scoring mechanism integrating risk and protective indicators. Advanced feature engineering produced composite indices—Parental Engagement, Student Commitment, Resource Availability, and Stress-Anxiety—while class imbalance was handled using SMOTE. Several feature-selection methods were evaluated, with LASSO proving most effective. Two ensemble models, Random Forest (RF) and XGBoost (XGB), were compared. Results show that XGBoost outperformed RF, achieving a higher F1-score and better capturing nonlinear relationships, while RF provided stable but lower performance. LIME-based interpretability indicated that high anxiety and reliance on academic support increased dropout risk, whereas strong resources and student commitment supported continuation. These findings highlight the importance of parental engagement, mental-health support, and socio-economic assistance to improve higher-education retention.