Learning Analytics based on Bayesian Optimization of Support Vector Machines with Application to Student Success Prediction in Mathematics Course

Abstract

Learning analytics is receiving a growing attention from both machine learning and education communities, where support vector machines (SVM) are gaining popularity over existing data mining techniques. In the scope of this work, we employ SVM to predict student success in mathematics course in Portugal under two common nonlinear kernel functions: polynomial and radial basis function kernel. In addition, we employ the k-nearest-neighbor (kNN) algorithm as a reference model since it is known to be fast and effective in various classification problems. Furthermore, we adopt the Bayesian optimization (BO) technique in a cross-validation framework to optimize SVM key parameters; namely, the slack parameter and penalty coefficient. The obtained experimental results show that the SVM outperform k-nearest-neighbor algorithm under both nonlinear kernel functions. Additionally, processing time associated with SVM optimization process increases with polynomial order. Furthermore, the SVM trained with third-order polynomial kernel performs the best. Finally, k-nearest-neighbor algorithm is found to be faster compared to all SVM classifiers.