Classifying Student Academic Achievement from Limited Categorical Institutional Records: A Comparative Study of Naive Bayes, K-Nearest Neighbor, and Decision Tree

Abstract

Student academic achievement prediction is an important application in Educational Data Mining (EDM) that supports proactive academic decision-making. This study investigates a specific and underexplored condition in the literature the classification of student academic achievement when the only available predictors are categorical institutional background attributes  without behavioral, attendance, or course-level data. This condition reflects data infrastructure limitations commonly found in Indonesian private higher education institutions. Three widely used classification algorithms Naive Bayes (BernoulliNB), K-Nearest Neighbor (KNN), and Decision Tree (CART) are compared against a majority class baseline through a five-stage preprocessing pipeline encompassing label normalization, cohort feature extraction, KNN k-value sensitivity analysis, and reporting of balanced accuracy and macro F1-score for fair evaluation under mild class imbalance. Results show that Decision Tree (depth=5) achieved the highest balanced accuracy (57.77%) and macro F1-score (57.51%), while Naive Bayes demonstrated the best generalization stability based on 10-fold cross-validation (60.07% ± 6.02%). All three models substantially outperformed the majority class baseline on balanced accuracy (+5–8 percentage points) and macro F1-score (+19–21 percentage points). Feature importance analysis identified IPS prior major background (15.6%) and the 2020 cohort (14.4%) as the most discriminative features. These findings provide evidence based algorithm selection guidance for data-constrained institutions and establish a reproducible performance benchmark for the categorical attributes only classification condition.

GS Cited Analysis