Dual Denoising Autoencoder Features for Imbalance Classification Problems
Ting Wang, Guangjun Zeng, Wing W. Y. Ng, Jinde Li · 2017
In pattern classification problems, it is difficult to force all classes to have the same number of training samples. Undersampling-based methods loss information while oversampling-based methods easily overfit. Therefore, the Dual Autoencoders Features (DAF) relieves the imbalanced pattern classification problem via learning a set of better features to project samples in different classes onto a more distinguishable space. However, the feature set learned by the DAF may not be robust to partial corruption of input patterns. Therefore, this work proposes a Dual Denoising Autoencoders Features (DDAF) to learn a more robust set of features to contaminated or destroyed training datasets. Experimental results show that the DDAF outperforms existing resampling-based methods and the DAF for imbalanced pattern classification problems.