Personalized Student Performance Prediction Using Multivariate Long Short-Term Memory

Abstract

In the age of knowledge economy, with deep international integration in all fields, the requirement for competitive human resources becomes increasingly fierce, deciding the success or failure of a country. The competitiveness of human resources depends much on the training process of the education system, notably the higher education level. Therefore, managing and monitoring student learning results are essential for lecturers, particularly the school in general. Early forecasting of students’ learning results is expected to help students choose the suitable modules or courses for their competencies, allowing leaders, administrators, and lecturers in universities and institutes to identify students who need more support to complete their studies successfully. In addition, it contributes to reducing academic warnings or expulsion or suspension due to poor academic performance. Also, this saves time and costs for students, families, schools, and society. This article proposes an approach to enhance student learning performance prediction by applying some deep learning techniques to exploit databases in student management systems at universities in a personalized way. More specifically, we consider personalized training, which means all mark entries of each student are trained separately and apply that trained model to predict scores of courses for themselves. The collected data is analyzed, pre-processed, designed, and prepared with a Long Short-Term Memory network with multiple input variables. Experimental results reveal that the proposed method’s average performance outperforms the method that trains whole datasets with an RMSE of 0.461 with a Multivariate Long Short-Term Memory network.