Comparison of Naïve Bayes, SVM, and Random Forest in Predicting the Job Waiting Time of Universitas Royal Alumni

Abstract

Alumni job waiting time is a graduate-quality indicator in tracer studies, yet its use for
prediction remains limited. This study compares Naïve Bayes, Support Vector Machine
(SVM), and Random Forest in predicting the job waiting time of Universitas Royal alumni.
The data come from the 2023–2025 tracer-study cohorts (598 alumni) with three waiting-time
classes (0 months, 1–3 months, >3 months) and two feature scenarios: basic and extended
(adding work region and job position). Processing was done in Google Colab using nested 5-
fold cross-validation repeated 5 times and Friedman and Wilcoxon tests. With basic features
the three algorithms were equivalent (macro-F1 0.373–0.383; p = 0.59). With extended
features Random Forest was highest (macro-F1 0.412), followed by SVM (0.400) and Naïve
Bayes (0.388), but the Friedman test was not significant (p = 0.14) and on the hold-out set
SVM was slightly higher. Test-set accuracy (scenario 1/2): Naïve Bayes 49.17%/50.83%,
SVM 47.50%/50.83%, Random Forest 50.00%/50.83%; baseline 54.17%. All models
exceeded the majority-class baseline on macro-F1 (0.234). Job position and employer type
were the most influential features. The findings are preliminary because the data are still a
sample.