Predicting Student Academic Performance in Video-Conferencing-Based Online Learning Environments

Abstract

Online learning, particularly video-conferencing has increased access to higher education, yet it is difficult to detect students at risk of compromised academic performance. In this paper, the relationships between online learning features and self-reported academic performance were examined, and five supervised machine learning classifiers were compared. The dataset of 29 variables and 361 records of a publicly available COVID-19-era questionnaire was analyzed; one of these records was excluded as a duplicate, leaving 360 observations. Three possible outcome-overlapping predictors were eliminated, which reduced the number of categorical predictors to 25. Evaluations of Logistic Regression, Decision tree, random forest, support vector machine and XGBoost have been done with 80: 20 stratified split and five-fold cross-validation. The results were good (72.5%), stable (22.8%), and decreased (4.7) performance. The highest reported relationship with performance was observed with ease of group work (Cramer's V = 0.194). The percentage of cross-validated macro-F1 (0.415) was highest in Random Forest and the test accuracy (73.6%), as well as macro-F1 (0.463) was highest in Support Vector Machine. The chosen Random Forest detected none of the three cases of decreased performance in the test subset, although it was 72.2% overall accurate. The permutation analysis showed significant and uncertain effects of learning ability and perceived efficiency. The results favor a careful exploratory application of prediction models but point to imbalance of classes, self-reported results, and poor minority-class detection. These models require larger externally validated datasets prior to making student support decisions.