RAGAN: Regression Attention Generative Adversarial Networks
Xiaoyu Jiang, Zhiqiang Ge
Abstract
Despite surrounding by Big Data, we still need to learn from insufficient data in many scenarios. Building an accurate regression model for a small amount of data is a pretty tricky and exciting problem. At present, it is a promising solution to augment limited real data by generating data through generative adversarial networks (GANs). However, when GAN is used to generate labeled data in regression modeling, it lacks attention to the relationship between independent and dependent variables, resulting in poor performance of regression modeling. This article proposes a novel regression attention GAN (RA-GAN) for augmented regression modeling. Regression attention mechanisms are introduced into network parameters learning of both generator and discriminator in RA-GAN to establish a known relationship between variables. This makes RA-GAN restore the regression information during data generation. In addition, an indicator called cross regression score is designed to describe the quality of the generated data before augmented regression modeling, effectively evaluating data augmentation performance in advance. The effectiveness and superiority of the proposed methods are verified in an actual industrial soft-sensing case and a diabetes prediction case through data augmentation regression applications.