Implementation of the C4.5 Decision Tree Algorithm in a Decision Support System for Employee Payroll at PT Hesed

Abstract

Payroll management is a critical business process because it directly affects employee compensation and organizational administration. PT Hesed still encounters challenges in payroll verification, including calculation errors, limited visibility of payroll components, and sequential verification procedures. This study implements the C4.5 Decision Tree algorithm in a web-based payroll decision support system and evaluates both software functionality and classification performance. The study employed a design and creation strategy using qualitative analysis for system requirement identification and quantitative analysis for model evaluation. Historical payroll data consisted of 12 verified records from January to April 2026, with payroll period, attendance, years of service, employment status, and tardiness as predictor attributes and payroll recommendation as the class label. C4.5 attribute selection used entropy, information gain, split information, and gain ratio, while Leave-One-Out Cross-Validation (LOOCV) evaluated predictive performance. Tardiness achieved the highest gain ratio (1.0000) and became the root node in the full dataset and in every LOOCV fold. LOOCV produced 7 correctly classified Appropriate records and 5 correctly classified Inappropriate records, yielding 100% accuracy, macro-precision, macro-recall, and macro-F1, compared with a 58.33% majority-class baseline accuracy. Black-box testing recorded outputs that matched the specified functions, while logic-level white-box verification reproduced the reference C4.5 calculations. Because the dataset is limited and no processing time, usability, payroll error reduction, or field-impact study was conducted, these results should be interpreted as evidence of functional conformity and internal classification consistency rather than demonstrated operational efficiency or broad generalizability.