Integrating GAN and sequential estimation for unbiased coefficient estimation and enhanced classification sensitivity in class-imbalanced applications

Abstract

Class-imbalanced data are common in biomedical and public health applications and often affect statistical inference and predictive performance. However, synthetic samples generated by traditional data-level methods raise concerns regarding their representativeness, making it questionable whether they truly originate from the target source population. Meanwhile, among algorithm-level approaches, some complex models lack interpretability, while others underestimate variance, leading to inadequate confidence interval coverage and biased inferences about feature effects. To address these challenges, we propose a framework that integrates Generative Adversarial Networks (GANs) with sequential estimation to handle class imbalance. First, we construct a geometry-preserving generative model based on Wasserstein GAN with Spectral Normalization (WGAN-SN) to generate high-fidelity minority samples. Second, we use sequential estimation and D-optimal design to select informative samples for building the logistic regression model, combined with Adaptive Shrinkage Estimation (ASE) for robust feature selection to ensure unbiased coefficient estimators. Simulation studies and real-world applications show that the proposed framework performs better and is practical to implement. Compared with standard methods, it yields unbiased coefficient estimates, provides reliable confidence interval coverage, and significantly improves minority-class identification sensitivity.

線上視訊請點選連結