COMPARATIVE ANALYSIS OF MACHINE LEARNING AND DEEP LEARNING APPROACHES FOR PREDICTING STUDENT DROPOUT IN HIGHER EDUCATION: A CASE STUDY OF THE VIRTUAL UNIVERSITY OF IVORY COAST

Student dropout is a critical challenge for academic governance and institutional performance in higher education systems. This research addresses the research: 'Which predictive models enable early identification of at-risk students, and what variables constitute relevant signals in this phenomenon?' Using institutional data from the Virtual University of Ivory Coast (UVCI), we develop and compare nine predictive models: five traditional machine-learning algorithms (Random Forest, Gradient Boosting, Support Vector Machines, Logistic Regression, Naive Bayes) and four deep learning architectures (Neural Networks, Deep Neural Networks, Transformer-based models, and a Hybrid Ensemble). The dataset comprised 9,881 student records with 14 features, preprocessed through null column detection and text normalization. We have rigorously defined the dropout prediction problem as a mathematical formulation through binary classification with class imbalance correction. There are five major predictive variables: final grade average (? = 0.847), number of uncompleted courses (? = 0.623), course completion rate (? = 0.591), course failure rate (? = 0.438), and student age (? = 0.216). Deep learning methods outperform other approaches in a comparative evaluation using the precision, F1-score, and AUC-ROC metrics. The best performance using Neural Networks is F1 = 0.9888 and accuracy = 0.9930, in comparison with the best machine learning model, namely Gradient Boosting: F1 = 0.9406, accuracy = 0.9641. Our mathematical modeling presents a rigorous foundation for an early warning system based on deep learning architectures, which will support targeted interventions and dynamic adaptation of learning pathways.