OAGAN: AN OVERSAMPLING APPROACH FOR IMBALANCED DATA PROBLEMS

Imbalanced data problems are prevalent in our daily lives. When conducted on this kind of data, many classification models may encounter many difficulties and produce ineffective performance due to a high training bias towards the majority of class instances. In order to solve the imbalanced data problem, many effective oversampling methods have been proposed. However, few methods in previous work consider different weights for minority class samples. Therefore, we proposed an oversampling method for the minority class by Adaptive Synthetic Sampling with Conditional Tabular Generative Adversarial Network (OAGAN) in this study. Firstly, Adaptive Synthetic Sampling is used to assign different weights to each minority class sample, based on which new similar samples are generated by the Conditional Tabular Generative Adversarial Network. In other words, the weights of different minority class samples are considered when generating data using the Conditional Tabular Generative Adversarial Network, thus improving the method. The proposed OAGAN method was evaluated on ten imbalanced datasets and three standard oversampling algorithms, which achieved better learning performance.