Student Performance Prediction Using Machine Learning Classification Models

The imperative to predict student performance has become a critical focus in educational research, serving as a powerful tool to enable early academic interventions and subsequently improve educational outcomes by addressing issues like academic failure and dropout. This paper addresses this need by investigating and comparing several machine learning classification models for predicting final student performance using the publicly available “Student Performance—Portugal Dataset”. This dataset is particularly valuable because it contains a rich set of features, including academic, demographic, and socio-behavioral factors (such as study time, family background, and alcohol consumption), which are crucial for modeling risk factors affecting achievement. We implement and evaluate three distinct supervised classification techniques: Logistic Regression, which serves as a simple and interpretable baseline model for categorical outcomes; the Random Forest ensemble method, known for its high accuracy and robust feature importance analysis; and the Support Vector Machine (SVM), which performs effectively on complex data structures using kernel-based decisions. Crucially, the final student grade (G3) is converted into a three-class prediction problem—Low ($0-9$), Medium ($10-14$), and High ($15-20$) achievement groups7777. This multi-class approach extends upon previous studies often limited to binary (Pass/Fail) classification, providing greater practical utility for academic decision-making and resource allocation8. Model performance is rigorously evaluated using a comprehensive suite of metrics: Accuracy, Precision, Recall, and the F1-macro score999999999. The Macro-average was specifically chosen to ensure all three performance classes are treated equally, providing a balanced assessment of predictive capability across the Low, Medium, and High groups10. Experimental results demonstrate that the Random Forest classifier achieves the best overall performance, exhibiting the highest accuracy and the strongest balanced predictive capability (F1-macro score of 0.70). Furthermore, the feature importance analysis reveals that study time and past failures are the most essential indicators of final performance. These findings collectively affirm the effectiveness of machine learning in educational analytics, providing actionable, data-driven insights that can significantly support decision-making in schools and universities.