Purpose: Wage-reporting fraud is a major challenge in social insurance systems. With the growing volume of data and the expansion of online premium‑submission systems, detecting such fraud through traditional methods has become increasingly difficult. This study aims to evaluate the effectiveness of unsupervised machine learning algorithms in identifying anomalies associated with wage-reporting fraud and to propose an automated approach for strengthening supervisory processes within the Social Security Organization.
Method: This applied research adopts a descriptive–analytical approach. Three unsupervised machine learning algorithms—Isolation Forest, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and One-Class Support Vector Machine (One-Class SVM)—were employed to detect anomalous wage records potentially indicative of fraud. The dataset consisted of 26,258 monthly wage records from 470 insured individuals covering the period from 2019 to 2023. Analyses focused on identifying abnormal patterns at both individual and cross-individual levels, and the performance of the algorithms was evaluated based on their detection logic and consistency with actual data behavior.
Findings: Isolation Forest demonstrated the most accurate and stable performance, identifying anomalies with a more reasonable and interpretable distribution compared with the other methods. DBSCAN exhibited excessive exclusion of observations in sparse data environments, while One-Class SVM showed high sensitivity accompanied by a higher rate of false alarms.
Conclusion: Unsupervised machine learning techniques provide an effective means for the automated detection of suspicious wage-reporting behaviors. The findings suggest that Isolation Forest can serve as a scalable and reliable solution for mitigating fraud risks in social insurance systems. It is recommended that this algorithm be utilized as the core component of an intelligent early-warning system within supervisory and monitoring frameworks.