Generator Information Enhancement Generative Adversarial Networks for Alleviating Data Imbalance Problems

In real-world datasets, numerous class distributions are imbalanced, which severely affects the classification ability of deep learning methods. Data augmentation can address such problem by creating data for the minority class to rebalance the datasets. However, it is still challenging that how to generate data from a true distribution of given finite training data. To overcome this challenge, we propose a novel and efficient synthetic oversampling approach, Generator Information Enhancement Generative Adversarial Networks (GIE-GAN). Its generative information is enhanced in two aspects. First, the generator can get common knowledge of all classes by initializing with a variational autoencoder. Second, during the training process, an independent classifier, which can provide accurate classification results for the generator, is added to the GAN for further augmenting the generative information. Experimental results on benchmark datasets show that, compared with the rival methods, our method can generate high-quality images and obtain better classifier performance of deep learning, especially for high imbalance ratios.

I. INTRODUCTION

The data imbalance is a widely existed and challenging problem in numerous real-world applications, such as medical diagnosis [1], fraud detection [2], fault detection [3] and others. The performance of most classification algorithms may significantly deteriorate under imbalanced data conditions, which results from the bias in the majority data class [4]. The common strategy to mitigate problems of this type is to resample the data before training models, including oversampling, under-sampling, and some kinds of combination of both. Among them, the work in paper [5] verifies that the oversampling techniques is the most effective way to handle the class imbalance for the Convolutional Neural Networks (CNN) model on image classification.