Medicare fraud detection: a comparative study on the effectiveness of one-class and binary classification models
Joffrey L. Leevy, John M. Hancock, Taghi M. Khoshgoftaar · International Journal of Computers and Applications · 2024
Research into Medicare fraud detection that utilizes machine learning methodologies is of great national interest due to the significant fiscal ramifications of this type of fraud. Our big data analysis pertains to three Medicare datasets: Part B, Part D, and DMEPOS. In this paper, One-Class Classification (OCC) and binary classification approaches are evaluated with six different classifiers. We note that when a practitioner is faced with a high cost or unreasonable wait associated with acquiring a second label for binary classification, OCC may intuitively seem like the better alternative. The evaluative metrics used in our study are Area Under the Receiver Operating Characteristic Curve (AUC) and Area Under the Precision-Recall Curve (AUPRC). As a result of our investigation, binary classification was found to be more effective than OCC in detecting Medicare fraud. Additionally, we determined that gradient-boosting algorithms, specifically CatBoost and XGBoost, yield the most robust performance.