Kevin Mills
Prediction of academic success has emerged as a crucial aspect of learning analytics and intelligent educational decision support. Much of the work done to date is still model-centric, and little is done to consider data quality, behavioural meaning, leakage prevention, explainability and the usefulness of interventions. This study introduces a data-centric Artificial Intelligence pipeline to predict the students' academic risk by analysing learning behaviour patterns. The Open University Learning Analytics Dataset (OULAD) was used, comprising records on demographics, registrations, assessments, Courses, virtual learning environments, and outcomes. The resulting analytical dataset consisted of 63 predictive features and 32,593 student observations, which had been preprocessed and engineered for the analysis. Several machine learning models were compared: Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbours, Naïve Bayes, Gradient Boosting, XGBoost, LightGBM, CatBoost and a neural network. The Random Forest model achieved the strongest F1-score of 0.8852, with 0.8840 accuracy, 0.9272 precision, 0.8469 recall, 0.9482 ROC-AUC, and 0.9625 PR-AUC. The proposed framework translates the predicted risk probability into levels of low, moderate, high and critical, to facilitate timely academic advising, tutoring, attendance review and retention planning. In general, the study illustrates the potential of the data-oriented approach to educational decision-making and intervention, offering greater explanatory value. Overall, the study shows the power of the data-oriented approach to educational decision-making and intervention that is more explanatory.