Research On Adversarial Classification Methods For Imbalanced Data

Class imbalance is a challenging and inevitable problem in a lot of real-world applications,such as medical imaging disease detection,financial risk assessment or text classification.In machine learning and deep learning,traditional classifiers always ignore the minority classes to ensure the overall prediction accuracy,resulting in poor performance of the minority classes.Unfortunately,misclassifying minority class instances often leads to heavy costs.In order to solve this problem,a simple and effective method is to rebalance the imbalanced data by augmenting minority class instances in the data level.The goal of this method is to generate new instances(especially minority class instances)with clear categories and sufficient intra-class diversity,which are really useful to the construction of classifier.However,it still suffers from two key issues.Firstly,data augmentation and classifier construction are performed separately,where classifier construction may not benefit from the augmentation strategies.Secondly,low variations in generated instances may lead to overfitting problem.In this paper,an end-to-end adversarial imbalance classification method based on Generative Adversarial Nets is proposed.It is a framework that unifies data augmentation and classifier construction,where these two stages enforce each other seamlessly.At the same time,it ensures sufficient intraclass diversity of generated instances.In order to solve the first issue,the class-specific and end-to-end adversarial imbalance classification strategy is proposed.This strategy draws on the idea of generative adversarial network.And it realizes the two stages compete and enforce each other,so that the generated instances are really beneficial to the construction of the classifier.In the meanwhile,the specially designed adversarial classifier drives the generation of class-specific instances by class-specific adversarial process so as to achieve the purpose of augmenting minority class instances.The experimental results demonstrate that the strategy effectively generates instances that are really beneficial to the construction of the classifier and improves the imbalanced classification performance in terms of three widely-used evaluation metrics on four benchmark datasets.In order to solve the second issue,the adversarial imbalance classification strategy based diversity gennerated instances is proposed.This strategy further assumes that the lattent space distribution is a Gaussian Mixture containing the distribution of the class during data augmentation.On this basis,a regularization constraint is added to the class distribution to enhance the intra-class diversity of generated instances.The experimental results show that the strategy can effectively improve the generalization performance of the classifier.