A Systematic Literature Review of a Machine Learning Model for Predicting Student Performance and Engagement in Higher Education

Abstract

The increasing availability of educational data has accelerated the application of machine learning (ML) to predict student outcomes and support data-driven educational decision-making. However, existing research varies considerably in algorithms, evaluation approaches, outcome variables, and geographical contexts. This study systematically analysed 17 research articles published between 2021 and 2025 that applied ML to predict student academic performance and engagement. The studies were examined in terms of prediction objectives, algorithms, reported performance, methodological approaches, and geographical contexts to identify dominant trends and research gaps. Random Forest (RF) emerged as the most frequently employed algorithm, appearing in 15 studies and frequently demonstrating strong reported performance. However, RF did not consistently outperform alternative algorithms. Some studies reported predictive accuracies approaching 100%; however, comparisons are constrained by differences in datasets, predictor variables, outcome definitions, validation procedures, and evaluation metrics. Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) were also frequently used. Academic performance was the predominant research focus (71%), while student engagement accounted for 29% of studies. The literature was geographically concentrated in Asia, with supervised learning approaches dominating. Emerging hybrid approaches combined clustering techniques, including Gaussian Mixture Models (GMM) and DBSCAN, with predictive modelling for student profiling. A significant gap was the limited integration of Explainable Artificial Intelligence (XAI), as most studies prioritised predictive accuracy over interpretability and actionable educational interventions. The findings indicate that while classical and ensemble ML models provide promising predictive capabilities, methodological heterogeneity, geographical concentration, and limited explainability constrain their practical utility. Future research should standardise evaluation practices, strengthen XAI integration, and expand research on engagement prediction and across diverse educational contexts to support systems linking prediction, explanation, and intervention.