Detecção de fraudes em auditoria interna por meio de algoritmos de aprendizado semi-supervisionado

Abstract

This study investigated the application of machine learning algorithms to support fraud and anomaly detection in accounting data from a logistics company. The dataset was extracted from the ERP (Enterprise Resource Planning) system, covering road fleet service records acquired between 2023 and 2025. Three semi-supervised learning strategies were compared, Active Learning, Pseudo-Labeling, and Label Propagation, applied to different models, including Logistic Regression, MLP, XGBoost, and CatBoost. The experimental protocol consisted of running each model-strategy combination over 10 independent repetitions, with stratified sampling and a fixed test set used exclusively for final evaluation. Results showed that Active Learning, particularly when combined with CatBoost, achieved the best performance, with an F1 score of 0.94 and PR-AUC of 0.98. As the second-best approach, CatBoost with Label Propagation reached an F1 score of 0.80 and PR-AUC of 0.87. In contrast, Pseudo-Labeling exhibited limitations in later iterations due to noise propagation from pseudo-labels. In practical terms, the findings demonstrate the potential of integrating semi-supervised learning techniques into internal auditing processes, enabling greater efficiency in prioritizing suspicious cases, reducing operational costs, and strengthening corporate governance.