Research On Imbalanced Data Based On GAN

Nowadays,we are in the era of big data,but the proportion of valuable data is relatively small.In practical application scenarios,there is often an unbalanced data distribution.If modeling with unbalanced data without processing,the trained models are often biased and cannot effectively identify the few categories of data.In the extreme imbalance case,the trained model is even invalid.We propose a variant of GAN,TOGAN(Target Oriented Generative Adversarial Networks),which predicts and solves the classification problem under imbalanced data.First,sort out the existing solutions imbalanced data and the basic principles of GAN and other generative models,and analyze the shortcomings of the existing methods.Then,on this basis,the basic framework of TOGAN is proposed,the necessity of the improvement of the method and the convergence of the algorithm are proved theoretically,and the simulation data is used to verify that it can be classified into two categories and multi-categories.Finally,an empirical analysis is done to apply TOGAN to image classification and text style transfer,which shows that TOGAN is suitable for a variety of data types and has excellent performance in practical applications.The main contribution and innovation of this thesis is to propose a classification model TOGAN for solving the imbalance of data distribution,and to demonstrate theoretically that TOGAN replaces the JSD of the original GAN with Wasserstein distance,which can alleviate the problem of gradient disappearance during training,and to verify this through simulations of binary and multi-classification.In addition,the simulation experiments show that TOGAN is more suitable for complex data distributions,and the data generated by TOGAN exhibit some diversity.In the empirical study,TOGAN achieved excellent results in both the image classification task and the text style transfer task.The image classification task uses the classical image classification dataset,the MNIST handwritten digit dataset,and the MNIST-FASHION dataset.The accuracy and recall of TOGAN classification are compared with the classification results of the original data,CVAE,and CGAN generated data,and TOGAN shows the most significant improvement in recall and an increase in accuracy.To verify the feasibility of TOGAN on other tasks,TOGAN is also modified and applied to the text style transfer task in this thesis.The GYAFC dataset is used for the text style transfer task,and the BLEU,GLEU,accuracy of TOGAN are compared with SimpleCopy,NMT,and DualRL,showing that TOGAN can significantly improve the style migration effect while maintaining the semantic content.