Abstract
Institutions of higher learning are currently facing the challenging task of attracting new students who can effectively meet their diverse academic demands. With these demands come the need for those institutions to develop strategies that can enhance students' learning experiences at various educational levels. Predicting the academic success at an early stage would allow academic institutions to develop specific enrolment guidelines while avoiding poor performance. The main purpose of this study was to predict the academic performance of students, their cumulative grade point average (CGPA) in particular, at postgraduate levels (e.g., master’s degree), using and comparing different machine learning (ML) algorithms. This work uses a real dataset of 635 master’s students collected from the college of graduate studies of a reputable private university in Malaysia. The predictive model’s goodness-of-fitness is determined using the coefficient of determination R2, which indicates the percentage of the variance in the dependent variables. The mean square error (MSE) and mean absolute error (MAE) are used to evaluate the model’s performance, by identifying discrepancies between the predicted CGPA and the actual CGPA. Of the six different ML models applied, our results showed that the artificial neural network (ANN) model has the best performance, achieving 89% of the variation in the CGPA of the students, with a training error of only 0.06 CGPA points and a prediction error of 0.08 CGPA points. The Gaussian process regression (GPR) model with squared exponential kernel algorithm achieved 71% of the CGPA variation. The model achieved 0.095 CGPA points for both training and evaluation errors. Exploring other variables such as research activities, marital status, and living conditions would have improved the overall accuracy of the proposed ML models. Therefore, future investigations should focus on predicting the academic performance of larger numbers of postgraduates (i.e., PhD and Masters) using different predictive variables and AI models.