Machine Learning for Predicting Student Academic Performance: A Case Study at a Public University

Abstract

Predicting how well a student will perform before, rather than after, a semester goes wrong is one of the more practical promises that machine learning has brought to higher education. This study reports a case study conducted at a public university, in which academic records, attendance logs, and a small set of socio-demographic indicators were used to build classification models that sort students into low, average, and high performance categories well before final examinations. Seven algorithms were trained and compared on the same dataset: logistic regression, decision tree, k-nearest neighbors, support vector machine, random forest, extreme gradient boosting (XGBoost), and a multilayer perceptron artificial neural network. After cleaning the data, encoding categorical fields, and selecting the ten most informative features through a correlation-based filter, the models were evaluated using accuracy, precision, recall, F1-score, and the area under the ROC curve, with a stratified 80/20 train-test split and five-fold cross-validation. XGBoost produced the strongest results, reaching 90.1% accuracy and an AUC of 0.95, followed closely by random forest and the neural network, while logistic regression and k-nearest neighbors trailed behind. Prior GPA, midterm score, and assignment performance emerged as the three most influential predictors, consistent with findings reported in comparable studies. The paper closes by discussing what these results mean for early-warning systems and academic advising at resource-constrained public universities, and where the approach still falls short. By turning data that public universities already collect into an early, actionable signal for advisors, this work speaks directly to Sustainable Development Goal 4 (Quality Education), particularly its emphasis on equitable learning opportunities and student retention.