A Study on the Application of the Secondary Anomaly Pattern Detection Model Based on Unsupervised Learning: Medicare Service Fraud Detection

Abstract

Subsidy is a system in which the government secures a budget for the social welfare of the people and executes it according to its purpose. Globally, the size of subsidies increases every year, and the types of subsidies are becoming various. However, contrary to the original good intention of subsidies, the scale and methods of fraudulent receiving are becoming more intelligent, so a countermeasure against this is urgently needed. Therefore, in this study, the second anomaly detection model developed to predict insurance fraud and unfair claims based on insurance industry data was extended and applied to the detection of subsidy fraudulent receiving patterns.

 For the fraud detection, this study applied the data from Medicare services similar to those of the U.S. health insurance service to the second anomaly detection model. Among the subsidies, the subsidy sector, which accounts for the largest portion of the budget, is the health and welfare sector, which also falls under Medicare. The model's core algorithm Isolation Forest, K-means, and Gaussian Mixture Model clustering algorithms were used to subdivide the characteristics of each group, and the binary classification of the unbalanced class was subdivided into four patterns and detected. Fraud patterns were detected by labeling fraud using U.S. Medicare service provider data and NPPES deactivation data provided by CMS, and classification performance was evaluated by learning pattern classification models for four patterns using machine learning and deep learning methods. Also, through pattern rule extraction, meaningful pattern rules were discovered by grasping the reliability of data classification according to each pattern.