Handling Class Imbalance Using Regularized Auto-Encoders with Weighted Calibration

Abstract

DeepSmote uses the SMOTE technique in the latent space of an Autoencoder- Decoder model to produce high fidelity images for imbalanced data. But it is be limited by 2 essential artillery: over-fitting the data and a lack of continuity of the latent space thus giving bad results. To overcome this, a number of regularized autoencoders have been proposed. Furthermore, the latent space was oversampled using a variety of approaches. Finally, a new method is a weighted calibration to the latent space of minority classes and has proven to be pretty accurate compared to other tested methods.