International Journal of Science, Engineering and TechnologyAuthors: Atul Sharma, Aashish Kumar Tiwari
Abstract: Identifying students who are likely to struggle or disengage before the problem becomes visible in a final grade is one of the more practical uses of machine learning in education. This paper presents a student performance prediction and early warning system that consumes academic records and learning-management-system interaction logs, engineers a feature set describing both prior achievement and in-course behavior, and trains several classifiers — logistic regression, decision tree, random forest, support vector machine, and a small feed-forward artificial neural network — to flag students at risk of poor performance. The models were trained and evaluated on a held-out split of a simulated cohort constructed to resemble a real learning-management-system export, and compared using accuracy, F1-score, and ROC-AUC. The random forest classifier reached the highest accuracy at 91.6% and an AUC of 0.94, ahead of the neural network and support vector machine, while logistic regression trailed the ensemble methods but remained useful as an interpretable baseline. The resulting risk scores are surfaced through an advisor-facing dashboard rather than shown directly to students, so that a human can decide how and when to intervene. The paper describes the system architecture, the feature engineering and modeling pipeline, the evaluation results with supporting figures, and the practical and ethical constraints — class imbalance, explainability, and the risk of self-fulfilling labels — that should guide any real deployment.
