Detecting Students’ Behavior on the E-Learning System Using SVM Kernels - Based Ensemble Learning Algorithm

Abstract

The corona pandemic has changed learning methods from face-to-face to online. However, the application of online learning creates difficulties for teachers in monitoring student behavior because of the reduced direct interaction. This problem causes the learning process to be less optimal. Moreover, students may fail to achieve learning objectives. This research addresses this problem by building a model to detect student behavior in this online learning. It focuses on finding an optimal model by exploring the ensemble learning-stacking method based on a combination of SVM kernels (Linear, Polynomial, RBF, Sigmoid). After the model was built, it was evaluated using two performance measurement techniques, namely: cross-validation and percentage split, and several performance measures, namely: AUC, Accuracy, F1, Precision, and Recall. The evaluation results show the superiority of the models applying ensemble learning over those without it. In terms of accuracy, the highest result in the cross-validation technique is 98.4%, achieved by three models employing stacking. Those three are with base learners combination of linear-polynomial-sigmoid kernel (LinPolSig_Stack), a combination of linear-RBF-sigmoid kernel (LinRBFSig_Stack), and a combination of all kernels-linear, polynomial, RBF, sigmoid (AllKernels_Stack). In the percentage split technique, the highest performance is 97.4%, achieved by two models implementing ensemble learning-stacking with base-learners combination of RBF-sigmoid kernel (RBFSig_Stack) and combination of linearpolynomial-sigmoid kernel (LinPolSig_Stack). Finally, the highest performance of these models is equivalent to the minimum error in detecting student behavior. Detection errors were only three students in the three models in the cross-validation technique and only six in the two models in the percentage split technique.