EARLY EDUCATIONAL PERFORMANCE PREDICTION A DEEP LEARNING APPROACH

Abstract

Early performance prediction is crucial for educators to identify struggling students. This
is especially important in a university where good students can perform badly due to multiple
external challenges. However, there are huge differences in terms of programs, policies as well
as culture between universities. These differences contribute significantly to students’
academic performance. Thus, it is important to address different universities separately to
predict students’ performance accurately. In this paper, an analysis of nearly 400 students’
records across 7 semesters of the same major in Hanoi University of Science and Technology is
presented. Because of the university privacy policy, it is impossible to obtain students
information other than their academic results. In addition, due to the modest size of the
datasets, imbalanced data is expected. Hence, we propose to use the Borderline SMOTE
algorithm to reduce the dataset’s imbalanced distribution. The data is then fed into a deep
neural network to predict students’ performance of the 4th year based on their previous years’
scores. A promising result of 77% accuracy is achieved.