Prediksi Kelulusan Tepat Waktu Mahasiswa Menggunakan Random Forest Pada Data Tidak Seimbang

Abstract

On-time graduation is an important indicator of higher education quality. This study applies the Random Forest algorithm to predict students’ graduation status based on six academic attributes from semester 1-4 (GPA semester 1, GPA semester 2, GPA semester 3, GPA semester 4, cumulative GPA, and completed credits), with Naïve Bayes as a comparison. The data were drawn from 1,020 students of the Informatics Engineering Study Program (2018-2021 cohorts). After data selection, 418 graduates were obtained, comprising 321 students (76.8%) who graduated on time and 97 students (23.2%) who did not. Since the data is imbalanced, three handling schemes were applied: default, class weighting, and SMOTE (Synthetic Minority Oversampling Technique). Evaluation was conduced using a single train-test split and 5-fold stratified cross-validation with accuracy, precision, recall, F1-score, and ROC-AUC. Results show that Random Forest with SMOTE achieved a ROC-AUC of 0.743, an average F1-score of 0.834 for the on-time class, and 50.6% ± 0.045 recall for the not on-time class, more stable than Naïve Bayes (F1-score 0.859, recall 44.5% ± 0.112). The Semester 3 GPA attribute was identified as the most influential predictor (23.3%). The results were implemented as a web-based application prototype using Streamlit as an early-detection.