Four Grade Levels-Based Models with Random Forest for Student Performance Prediction at a Multidisciplinary University

Abstract

Student performance is a critical task in universities. By predicting student performance in the early stage, we can identify students who need more attention to improve their learning performance. Also, these forecast tasks support students to select appropriate courses and design good study plans for themselves to obtain higher performance. Previous studies usually use only one model for all kinds of students regardless of each student’s ability and characteristics. For multidisciplinary universities, this type of model can produce poor performance. In this study, we propose to consider 4-grade levels of degree classification in Vietnam, and the prediction is based on the average performance in previous semesters to perform the prediction tasks. Four student groups are trained separately with their mark records—the used model depends on students’ average marks in previous semesters. The proposed method is validated on more than 4.5 million mark records of nearly 100,000 students at a multidisciplinary university in Vietnam. The experimental results show that the four Random Forest-based models give a positive average mean absolute error of 0.452 of Random Forest regression comparing with the error of 0.557 while using one model.