Improving Medicare Fraud Detection through Big Data Size Reduction Techniques

Classification models serve as effective tools for Medicare fraud detection, but their performance can be influenced by a number of factors. This paper focuses on addressing two common challenges using the Medicare Part D Big Data: high dimensionality and class imbalance. To tackle these issues, we explore the use of feature selection, Random Undersampling (RUS), and a combination of our novel feature selection technique followed by RUS. RUS is employed to create five different class ratios, while six supervised feature selection methods are utilized within ensemble feature selection techniques. The performance of six machine learning classifiers is evaluated for Medicare fraud detection, using Area Under the Receiver Operating Characteristic Curve (AUC) and Area Under the Precision-Recall Curve (AUPRC). Additionally, we compare the performance of models built with the original dataset to establish the superiority of our technique over a baseline approach. The results clearly demonstrate that reducing the training datasets significantly using RUS with a minority:majority class ratio of 1:81 and a feature subset size of ten leads to similar or improved model performance compared to models built with the original dataset in terms of AUPRC metric. Additionally, the data reduction allows for faster model training times. Therefore, details on how the results were obtained are a contribution to Big Data research.