A Leakage-Aware and Decision-Oriented Multi-Model Analytics Approach for Predictive Risk Management in Educational Leadership

Leadership education (LE) can strengthen student success when leaders act early on credible, interpretable signals of academic risk. This paper evaluates a leakage-aware machine learning pipeline for two early-warning tasks: absence-risk classification and performance-risk screening. The pipeline combines categorical encoding, feature scaling, in-fold SMOTE, PCA, stratified cross-validation, and model-specific hyperparameter tuning, and benchmarks five tabular classifiers: K-nearest neighbors (KNN), random forest (RF), gradient boosting (GB), XGBoost, and LightGBM. Two public educational datasets are used to test the pipeline across a multi-class target and a binary target with different sample sizes and imbalance profiles. Results are reported with accuracy, macro-averaged precision, recall, and F1-score, together with confusion-matrix interpretation. The study contributes a reproducible comparison protocol and a decision-oriented interpretation of model errors; it does not claim that the models are intrinsically fair or that leakage prevention is a mathematical guarantee. The findings show that model performance depends on the risk definition and dataset structure, supporting evidence-based model selection and human-governed intervention.