Combining Self-Training and Tri-Training for Course-Level Student Classification

Abstract

Course-level student classification is one of the most popular tasks in the educational data mining area. It helps early identifying the students facing difficulties in a course by predicting their final study status after the end of the course. To perform this task more effectively, our work proposes a semi-supervised learning method, TriForest, which combines Self-Training and Tri-Training. TriForest has three base classifiers all of which are Random Forests in the learning process of the Self-Training framework. Meanwhile, TriForest explores the agreement mechanism of Tri-Training in selecting the most confidently predicted instances for training data set enhancement of each base classifier. In addition, our work also takes into account both study performance via grades and forum-related attributes in this classification task. The selected attributes are useful for the proposed method to differentiate successful students from students with failure in a course where the data sets of the past students can be used to build a classification model for the current students. Indeed, experimental results on real data sets have shown that with about 90% of Accuracy, our method can achieve better predictions than its base classifier and other semi-supervised learning methods. Such accurate student classifications help both students and lecturers forecast the students' study status and in-trouble students can be then supported appropriately for their ultimate success in the course.