Cross-Fitted Multi-View Neural Risk Augmentation for Interpretable University Dropout Prediction

Abstract

Early dropout prediction requires models that integrate heterogeneous educational records without information leakage. We propose multi-view neural risk augmentation (MVNRA), a cross-fitted framework that separates early-warning predictors into academic, contextual, and digital-engagement views. View-specific encoders and a gated fusion module produce four supervised risk scores. We evaluated MVNRA on a public longitudinal dataset from a Spanish technological university containing 464,739 student–course records and 81 predictors available by December. Within each outer student-level GroupKFold split, the neural risk generator was trained through inner student-grouped cross-fitting. Its out-of-fold fused, academic, contextual, and digital risk scores were then appended to the original predictors before six conventional classifiers were fitted. MVNRA increased mean area under the receiver operating characteristic curve (AUC) for all six classifiers. The largest gains occurred for AdaBoost (0.7717 to 0.9240), linear discriminant analysis (0.8701 to 0.9176), Gaussian Naive Bayes (0.7313 to 0.7786), and Logistic Regression (0.9139 to 0.9352). MVNRA + Random Forest achieved the highest absolute discrimination (AUC =0.9577), with increases of 0.0191 in precision–recall AUC and 0.0112 in F1 score. SHAP analysis ranked the fused and academic neural risk scores above the cumulative academic-progress indicators. These findings show that cross-fitted, view-specific neural risk scores can transfer nonlinear structure to conventional tabular classifiers while preserving a SHAP-compatible decision layer and strict student-level separation.