The Use of Supervised and Unsupervised Learning Methods for Detecting Auditing Anomalies

Abstract

The rapid growth in transaction volumes and increasing sophistication of fraudulent schemes have reduced the effectiveness of traditional audit sampling in modern financial systems. This study evaluates supervised and unsupervised machine learning techniques for audit anomaly detection and compares their effectiveness across anomaly types and organizational contexts. Using a real-world dataset of 247,683 financial transactions, we implemented five supervised algorithms—Logistic Regression, Random Forest, Support Vector Machines, XGBoost, and Artificial Neural Networks—and five unsupervised methods—K-means, DBSCAN, Isolation Forest, Local Outlier Factor, and Autoencoders. Supervised models were evaluated using precision, recall, F1-score, and AUC-ROC, while unsupervised models were assessed using detection rate, false positive rate, anomaly scores, and silhouette coefficients. Results show that ensemble supervised techniques perform best when labeled data are available. XGBoost achieved the highest performance with an F1-score of 0.89 and AUC of 0.94, while maintaining a false positive rate below 1%. Among unsupervised approaches, Isolation Forest achieved an 87.3% detection rate with a 4.2% false positive rate, and autoencoders demonstrated competitive performance in high-dimensional anomaly detection. The findings indicate that supervised models are more accurate for detecting known fraud patterns, whereas unsupervised methods are essential for identifying emerging or previously unseen anomalies, particularly in environments with limited labeled data. The study supports the adoption of hybrid audit analytics frameworks that combine both paradigms to improve detection coverage while balancing interpretability, computational efficiency, and operational feasibility. These results provide practical guidance for auditors implementing machine learning–based anomaly detection syste