Comparative Analysis of Deep Learning and Gradient Boosting for Student Dropout Prediction on OULAD

Student dropout in e-learning remains a persistent problem, and the belief that deep learning outperforms conventional methods has rarely been tested on large, real data. This study compares five models: Random Forest, XGBoost, a convolutional neural network (CNN), a long short-term memory (LSTM) network, and a hybrid CNN-LSTM, for dropout prediction on the Open University Learning Analytics Dataset (OULAD), covering 32,593 students with 21 aggregated features shared by all models. The Synthetic Minority Over-sampling Technique (SMOTE) was applied within each fold of a 5-fold stratified cross-validation, and the best model was explained with SHapley Additive exPlanations (SHAP) and converted into a three-tier Early Warning System (EWS). XGBoost reached the highest accuracy (86.77%) and area under the receiver operating characteristic curve (AUC) of 0.9442, ahead of the hybrid CNN-LSTM (accuracy 85.36%, AUC 0.9328). The gap had a large effect size (Cohen's dz = -1.06) but was not significant (p = 0.077). SHAP identified active span as the strongest predictor, and the EWS flagged 10,180 students as high-risk. External validation on an independent public dataset reproduced the ranking (XGBoost AUC 0.923; deep models mean AUC 0.879). Interpretable gradient boosting therefore matches or exceeds deep learning on tabular data.