Distance Guided Generative Adversarial Network for Explainable Binary Classifications

TL;DR

Data scarcity in binary classification often leads to overfitting and opaque decision boundaries. The authors propose DisGAN, a distance-guided GAN that conditions inter-domain and intra-domain generation on vertical and horizontal distances to a fixed optimal hyperplane $C(z)=w^T z + b$, thereby reshaping the margin and improving explainability via class-difference maps. VerDisGAN and HorDisGAN, along with cycle-consistency and distance losses, demonstrate improved ACC and AUC on natural and medical datasets, and provide interpretable localization of decision-relevant regions. The framework is architecture-agnostic, scalable to multiple backbones, and extensible to multi-class settings, offering a practical path for robust, explainable data augmentation in limited-data scenarios.

Abstract

Despite the potential benefits of data augmentation for mitigating the data insufficiency, traditional augmentation methods primarily rely on the prior intra-domain knowledge. On the other hand, advanced generative adversarial networks (GANs) generate inter-domain samples with limited variety. These previous methods make limited contributions to describing the decision boundaries for binary classification. In this paper, we propose a distance guided GAN (DisGAN) which controls the variation degrees of generated samples in the hyperplane space. Specifically, we instantiate the idea of DisGAN by combining two ways. The first way is vertical distance GAN (VerDisGAN) where the inter-domain generation is conditioned on the vertical distances. The second way is horizontal distance GAN (HorDisGAN) where the intra-domain generation is conditioned on the horizontal distances. Furthermore, VerDisGAN can produce the class-specific regions by mapping the source images to the hyperplane. Experimental results show that DisGAN consistently outperforms the GAN-based augmentation methods with explainable binary classification. The proposed method can apply to different classification architectures and has potential to extend to multi-class classification.

Figures (8)

• Figure 1: The illustration of data augmentation for binary classification under two settings: (a) Not controlling the variation degrees. The generated samples are intra-domain by traditional augmentation; and degenerate to be intra-domain due to model collapse and have uncertain domain labels by GAN-based augmentation. (b) Controlling the variation degrees. The generated samples can reshape the decision boundary in the hyperplane space.
• Figure 2: Overview of the proposed data augmentation method. (a) Training a binary classifier to obtain an optimal hyperplane via hinge loss. The fixed classifier is taken as an auxiliary classifier for image generation. (b) The DisGAN consists of VerDisGAN and HorDisGAN. They map the source images to the target images by taking the distances as input parameters. The auxiliary classifier is used to reconstruct the input distances from the generated images. In addition, they map the generated images back to source images. The reconstructed source images should be consist with the source images.
• Figure 3: Qualitative results over Butterfly Mimics butterfly and Asian vs African Elephants GoumiriBP23. The generated images by the DisGAN are clearly more realistic than that by the CycleGAN cyclegan. The distance parameters for generating these images are provided by the auxiliary classifier ConvNeXt convnext.
• Figure 4: Qualitative results over mixed Breast Ultrasound BUSIUDIAT and COVID-CT zhao2020COVID-CT-Dataset. The generated images by the DisGAN is clearly more realistic than that by the CycleGAN cyclegan. The distance parameters for generating these images are provided by the auxiliary classifier ConvNeXt convnext.
• Figure 5: Comparing DisGAN's class-difference maps (CDMs) with Grad-CAM gradcam (TA + CrossEntroy) over the mixed Breast Ultrasound and the COVID-CT datasets. The difference between the source images and their projections on a hyperplane can highlight class-specific region, which cannot be deduced from the Grad-CAM gradcam.