A STACKING BASED HYBRID TECHNIQUE TO PREDICT STUDENT DROPOUT AT UNIVERSITIES

University dropout is a very complex problem that affects the Government, Institutions and students and families in the world. The prediction allows to identify the students who are going to desert early, so that the directors of the Universities can establish strategies to mitigate it. Machine Learning methods are the most recent and effective for this problem. However, so far these methods have been applied independently and not in combination. This paper proposes a hybrid model based on decision trees and neural networks, designed following the KDD methodology, to predict with high precision the university student dropout. The proposal was implemented in Rapid Miner Studio 6.4 and applied to a dataset with 1761 student records and 53 variables for training. Through a variable selection procedure that includes 8 algorithms, 27 variables were selected. The results on 100 new records show an accuracy of 87%, 91%, 98% for decision tree models, neural networks and stacking respectively. In addition, the result of sensitivity is 90.6%, 93.3%, 98.7% for decision trees, neural networks and stacking respectively. Regarding specificity, 76%, 84% and 96% have been obtained for decision trees, neural networks and stacking respectively. The results of accuracy, sensitivity and specificity also show that the hybrid model presents better results than the separate models..