Authors - Petros Papagiannis, George Pallaris, Pantelitsa Leonidou Abstract - Predicting student academic performance presents a persistent challenge for higher education institutions. This paper presents a machine learning study at Cyprus College, Cyprus, using 311 studentcourse records across three semesters from 13 Computer Science courses. Four assessment components—midterm examination, final examination, assignments, and participation—alongside absence counts for 95 unique students are used as features. Seven algorithms are evaluated—Random Forest, XGBoost, SVM, Logistic Regression, K-Nearest Neighbours, Decision Tree, and Naive Bayes—using stratified five-fold cross-validation across three tasks: regression, binary pass/fail classification, and multiclass grade band prediction. SHAP analysis (applied to the full feature set) identifies feature contributions, while early-warning experiments exclude the final examination score to simulate mid-semester prediction. Results show that midterm and assignment scores predict final outcomes with R2=0.746 before the final examination, whilst Random Forest and XGBoost achieve 97.4% pass/fail accuracy. Participation contributes zero predictive signal despite a 10% grade weighting, with direct implications for assessment design at small higher education institutions.
Open Zoom