Latent Posterior Based Generative Adversarial Network for Imbalance Classification

The imbalance classification problem arises when certain class is underrepresented in comparison with other classes, leading to a classifier partial to the majority classes. Existing interpolation-based oversampling methods for handling this problem characteristically do not make full use of the probability distribution of data. To overcome this weakness, this study proposes latent posterior based generative adversarial network oversampling approach(LPGOS), which uses a variational encoder to obtain the posterior distribution of latent variables. In addition, giving the high correlation between generated synthetic data and original data, we introduce a transfer learning approach with weight scaling factor namely TrWSBoost in which the generated minority class samples are treated as source domain data. Visual results prove that the proposed approach LPGOS is capable of approximate high dimensional data distribution and outperform other existing oversampling techniques. The performance of binary classifiers verify the effectiveness of proposed approaches.