Searching for Optimal Oversampling to Process Imbalanced Data: Generative Adversarial Networks and Synthetic Minority Over-Sampling Technique

Classification problems due to data imbalance occur in many fields and have long been studied the machine learning field. Many real-world datasets suffer from issue of class imbalance, which occurs when sizes classes are not uniform; thus, belonging minority likely be misclassified. It is particularly important overcome this dealing with medical because inevitably arises incidence rates within datasets. This study adjusted ratio (IR) National Biobank Korea dataset “Epidemiologic Parkinson’s disease dementia patients” values 6.8 (raw data), 9, 19 compared four traditional oversampling methods techniques using conditional generative adversarial network (CGAN) tabular (CTGAN). The results showed that were balanced CGAN CTGAN, they a better classification performance than more based on AUC F1-score. We able expand application scope GAN, widely used unstructured data, structured data. also offer solution for imbalanced problem suggest future research directions.