Comparison of Machine Learning Methods for Early Detection of Student Dropouts

One of the major issues in Educational Artificial Intelligence is detecting university student dropouts. Student dropout prediction is a growing research area since dropouts have financial, social, and national consequences. To deal with this issue, we examined the performance of ten machine learning algorithms with different feature sets in predicting university student dropout. Used algorithms are: Decision Trees, K-Nearest Neighbors, Naïve Bayes, Logistic Regression, Stacking Classifier, Adaboost, XGBoost, Random Forest, Support Vector Machines, and Multi Layer Perceptrons. The Synthetic Minority Oversampling Technique (SMOTE) balancing algorithm is used in the training process, to balance training data since the used dataset has an imbalanced nature. On the student dataset from Tecnologico de Monterrey in Mexico, we obtained 92 % accuracy on the XGBoost algorithm with a sub-feature set. We showed that the selected data features are as important as the selected method. Considering different performance metrics other than accuracy for imbalanced data is important. Having economic and historical score data increases accuracy. We also have seen that XGBoost and Random Forrest algorithms were the best for this task.