IEEE Access (Jan 2023)

Machine Learning in Higher Education: Students’ Performance Assessment Considering Online Activity Logs

  • Ghazanfar Latif,
  • Sherif E. Abdelhamid,
  • Khaled S. Fawagreh,
  • Ghassen Ben Brahim,
  • Runna Alghazo

DOI
https://doi.org/10.1109/ACCESS.2023.3287972
Journal volume & issue
Vol. 11
pp. 69586 – 69600

Abstract

Read online

Machine learning in Education is receiving more attention from researchers as the number of students at all levels globally is increasing. To ensure students’ success in K-12 educational institutions and higher education institutions work needs to be done to assist students, teachers/professors, parents, and all stakeholders to provide the support that students need. The need and motivation for such systems are very well-established and thus the aim of this work is to develop a system based on modified machine learning models to automatically predict students’ performance and subsequently identify students at risk. The DEEDs dataset is used in this study. Novel features were extracted and applied to well-known classifiers some of which are ensemble classifiers. These classifiers were also combined with base learners such as bagging and boosting. The problem was divided into three scenarios; binary classification of the pass and fail, three class scenarios, and four class scenarios. It was shown that ensemble methods combined with base learners of boosting and bagging significantly increase the accuracy for binary classification, slightly increase accuracy for three class problems, and have no significance in increasing the accuracy when the problem is 4 classes. The ensemble algorithm of bagging and boosting FDT achieved an accuracy of 98.25% for binary classification and 89.47% for three classes. The standard ensemble FDT achieved an accuracy of 77.19% for four classes. The results obtained for binary classification were compared with results reported in the extant literature using the same dataset proving that the proposed modified algorithms achieved better results than similarly proposed methods. The three-class and four-class results could not be compared because according to the author’s knowledge, there are no research papers published for the same dataset for multi-class classification.

Keywords