Predicting Student Achievement Status Using Decision Tree and Random Forest with SHAP in Vocational School Admissions

Abstract

This study investigates the ability of Decision Tree and Random Forest models to predict student achievement status using a synthetic vocational-school admission dataset. The dataset contains 1,000 records with Achievement as the target, consisting of 215 achievement and 785 non-achievement records. Predictors include academic year, school origin, vocational program, report-card score, Mathematics, Indonesian Language, English, selection-test score, and home distance. The data were divided using a stratified 80:20 train-test split, and both models were evaluated using accuracy, balanced accuracy, precision, recall, specificity, F1-score, ROC-AUC, MCC, and confusion matrices. Random Forest achieved 77.00% accuracy but only 51.58% balanced accuracy and 45.28% ROC-AUC, whereas Decision Tree obtained 57.00% accuracy, 46.44% balanced accuracy, and 46.05% ROC-AUC. SHAP analysis of Random Forest identified vocational program, Mathematics score, English score, and selection-test score as the strongest global contributors. The results demonstrate that high accuracy on imbalanced data does not necessarily indicate strong discriminative performance. The synthetic dataset does not contain sufficiently strong predictive signals for reliable achievement classification, highlighting the need for validation using real student-achievement data.