Predicting academic performance by considering student heterogeneity

• Published in: Knowledge-based systems 2018-12, Vol.161, p.134-146
• Main Authors: Helal, Sumyea, Li, Jiuyong, Liu, Lin, Ebrahimie, Esmaeil, Dawson, Shane, Murray, Duncan J., Long, Qi
• Format: Article
• Language: English
• Subjects: Academic achievement; Artificial intelligence; Colleges & universities; Data mining; Demographics; Distance learning; Educational data mining; Expert systems; Heterogeneity; Learning management system; Learning management systems; Mathematical models; Performance prediction; Student heterogeneity; Student performance prediction; Students; WEKA
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2018.07.042

The capacity to predict student academic outcomes is of value for any educational institution aiming to improve student performance and persistence. Based on the generated predictions, students identified as being at risk of academic retention or performance can be provided support in a more timely manner. This study creates different classification models for predicting student performance, using data collected from an Australian university. The data include student enrolment details as well as the activity data generated from the university learning management system (LMS). The enrolment data contain student information such as socio-demographic features, university admission basis (e.g. via entry exam or past experience) and attendance type (e.g. full-time vs. part-time). The LMS data record student engagement with their online learning activities. An important contribution of this study is the consideration of student heterogeneity in constructing the predictive models. This is based on the observation that students with different socio-demographic features or study modes may exhibit varying learning motivations. The experiments validated the hypothesis that the models trained with instances in student sub-populations outperform those constructed using all data instances. Furthermore, the experiments revealed that considering both enrolment and course activity features aids in identifying vulnerable students more precisely. The experiments determined that no individual method exhibits superior performance in all aspects. However, the rule-based and tree-based methods generate models with higher interpretability, making them more useful for designing effective student support.