Mike-Kirkpatrick/A-Comparative-Analysis-of-Machine-Learning-Techniques-for-Predicting-Student-First-Year-Dropout
ABSTRACT: Student dropout is of the utmost concern in higher education and machine learning techniques have become a powerful tool for proactively identifying students at-risk of dropping out. Data from more than 17,000 National University students were used to train nine machine learning algorithms to predict first-year dropout under four conditions, thus resulting in thirty-six models. The algorithms included were Logistic Regression, Naïve Bayes, Neural Networks, k-Nearest Neighbor, Support Vector Machine with linear and polynomial kernels, Decision Tree, Random Forest, and XGBoost. Modeling conditions varied with regard to class balancing and feature reduction. Models were evaluated based on ROC area and accuracy. Ensemble tree-methods XGBoost and Random Forest were superior across all modeling conditions. Overall, class balancing and feature reduction did not improve model performance. Feature importance was examined and many novel features proved to be useful for dropout prediction. Recommendations for