International Journal of Science, Engineering and TechnologyAuthors: Babandi Usman, Salim Ahmad, Zahraddeen Sufyan
Abstract: The early identification of students who are likely to experience academic failure has become an important application of machine learning (ML) and learning analytics in higher education. Although numerous studies have developed predictive models for student performance, the literature remains fragmented across performance prediction, at-risk identification, dropout prediction, and student success modelling, with substantial variation in outcome definitions, predictor availability, validation strategies, and evaluation metrics. This systematic literature review synthesizes research on ML-based prediction of student academic outcomes, with particular emphasis on failure-oriented and early-warning applications. A PRISMA-aligned evidence-synthesis strategy was used to examine systematic reviews, empirical studies, established educational datasets, and recent developments in explainable and responsible artificial intelligence. The reviewed evidence indicates that classification-based approaches, particularly decision trees, random forests, support vector machines, logistic regression, artificial neural networks, and ensemble methods, dominate the field. Academic history, assessment performance, learning-management-system activity, attendance, engagement, and demographic or contextual variables are frequently used predictors. However, high predictive accuracy does not necessarily indicate practical usefulness because information leakage, class imbalance, temporal instability, inadequate calibration, limited external validation, and weak interpretability remain important methodological concerns. Recent studies further demonstrate the importance of explainable AI, fairness assessment, privacy-preserving data generation, and temporal modelling. Based on these findings, this study proposes a Failure-Oriented Explainable Temporal Machine Learning Framework (FET-MLF), integrating leakage-aware temporal feature engineering, predictive modelling, probability calibration, explainability, fairness auditing, and intervention feedback. The framework shifts evaluation from accuracy-centric modelling toward trustworthy, actionable, and institutionally deployable early-warning systems for preventing academic failure.
