Feature Transformation Based on Autoencoder to Oversample on Imbalanced Data

Imbalanced learning has gained many attentions in the domain of data mining and machine learning. Traditional classifiers often suffer from the under-representation of minority data with insufficient quantity in the training process. The remarkable data reconstruction abilities exhibited by autoencoders (AE) present us with a novel approach to oversample, named AE-OM. In this approach, AE is firstly hired to learn a manifold mapping that compresses critical features of majority samples nonlinearly. Then the manifold mapping is operated on minority samples to reconstruct new minority samples, where minority samples travel the same information route. The reconstructed minority samples not only capture the distribution information of raw positive samples but also filter noise. Conditional classifiers trained on datasets balanced with proposed method, AE-OM shows superior performance in terms of metrics of G <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</inf> , F <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1,</inf> and AUC compared with 15 popular imbalanced methods.