Deep Learning based Ensemble Approach to Predict Student Academic Performance: Case Study

Traditional practices of education, medicine, and economics often use collective expert methods to find acceptable solutions to complex issues. The computer model of such methods is an ensemble classification which combines prediction of a number base classifier trained by various machine learning techniques. The predictive sample in the educational process is always not balanced - it leads to a shift in estimates. The basic idea of this work is that the binary forecasting on unbalanced datasets is carried out by a Meta-classifier - an ensemble of different classification algorithm (Decision tree J48, Naive base NB, Multilayer perception MLP, k- Neighbour Nearest k-NN and Support Vector Machine SVM) which trained by stacking, bagging and boosting procedures on SMOTE, RUS, ROS balanced samples. The experiments conducted on moderately unbalanced dataset of 350 tuples of student data with using simple voting Meta-ensemble models, show that the best overall accuracy A and F1 measure of minority class were achieved by the Meta-ensembles containing 5-15 members classifiers trained on samples prepared by SMOTE. It is also proposed that these of Meta-ensembles can be used to resolve quantification learning problem - novel task to predict collective performance. As in our recent works the deterministic outputs of Meta-ensemble members can be transformed to probabilistic outputs and then to prediction of class frequency. After analyzing the results of quantification, it becomes obvious that the most optimal design of the meta-ensembles is to collect pre-boosted classifiers trained on the SMOTE sample.