Sampling Approaches for Imbalanced Data Classification Problem in Machine Learning

Real-world datasets in many domains like medical, intrusion detection, fraud transactions and bioinformatics are highly imbalanced. In classification problems, imbalanced datasets negatively affect the accuracy of class predictions. This skewness can be handled either by oversamplingminority class examples or by undersampling majority class. In this work, popular methods of both categories have been evaluated for their capability of improving the imbalanced ratio of five highly imbalanced datasets from different application domains. Effect of balancing on classification results has been also investigated. It has been observed that adaptive synthetic oversampling approach can best improve the imbalance ratio as well as classification results. However, undersampling approaches gave better overall performance on all datasets.