Learning from multiple dynamic graphs of student and course interactions for student grade predictions

Abstract Predicting students’ course grades has important university-related applications, such as providing predictive assistance to students who may fail to graduate. However, it is a quite challenging task. On the one hand, each student has very limited historical course grades. On the other hand, a student may exhibit very different performance on different types of courses. As a result, most existing grade prediction methods don’t address such challenges and cannot achieve good results. Through empirical data analysis, we find that a group of students achieve similar grades over a set of similar courses. Based on the observations, we first construct student-course graphs, student-student graphs and course-course graphs to capture student-course dependency, student similarity and course similarity, respectively. Then, we propose a model named DGTEAM to specifically deal with these three dynamic graphs to obtain the representations of students and courses. The obtained representations of each course and student are applied to predict the grades. We conduct experiments on a real-world dataset and the result verifies the superiority of our model.