Distance Guided Generative Adversarial Network for Explainable Binary Classifications

Xiangyu Xiong, Yue Sun, Xiaohong Liu, Wei Ke, Chan-Tong Lam, Jiangang Chen, Mingfeng Jiang, Mingwei Wang, Hui Xie, Tong Tong, Qinquan Gao, Hao Chen, Tao TanarXiv 2023 · DOI · arXiv

Introduction

In modern deep learning, the performance of a classifier is often limited by the availability of high-quality, labeled data. This is particularly challenging in niche fields such as medical imaging or rare species identification, where datasets are inherently small. While traditional data augmentation techniques like rotation or flipping provide some relief, they do not introduce new feature-space information. Generative Adversarial Networks (GANs) offer a potential solution by synthesizing new samples; however, standard GANs often suffer from "mode collapse" or generate samples that are either too similar to existing data (intra-domain) or too ambiguous to help define a clear decision boundary (inter-domain).

Conceptual comparison between standard GANs (a) and DisGAN (b), illustrating how DisGAN fills the margins near the decision boundary.

To address these limitations, the Distance Guided Generative Adversarial Network (DisGAN) framework was developed. DisGAN moves beyond simple image generation by strategically "sculpting" the decision boundary. By generating samples with mathematical precision relative to a classifier's hyperplane, it ensures that synthesized data is informative and robust, directly improving classification accuracy and model interpretability.

Methodology: The Hyperplane and Distance Guidance

The core innovation of DisGAN lies in its use of a hyperplane space to guide the generative process. The framework operates in three distinct phases:

  1. Hyperplane Construction: A binary classifier is initially trained using a hinge loss function to identify an optimal linear hyperplane that separates the two target classes. This hyperplane serves as the reference point for all subsequent generation.
  2. Distance Measurement: The system calculates two critical metrics for every sample:
  3. Vertical Distance ($d_v$): The distance from a sample to the hyperplane, representing the confidence or strength of its class membership.
  4. Horizontal Distance ($d_h$): The distance between samples within the same domain, which serves as a proxy for diversity and variation.
  5. Dual-GAN Architecture: The framework employs two specialized generators:
  6. VerDisGAN (Vertical Distance GAN): Focuses on inter-domain generation (e.g., transforming Class A to Class B). It is conditioned on vertical distances to ensure generated samples land in specific, useful regions of the target domain.
  7. HorDisGAN (Horizontal Distance GAN): Focuses on intra-domain generation (e.g., Class A to a varied Class A'). It uses horizontal distances to ensure the generated samples are diverse and effectively cover the margins of the class distribution.
The DisGAN framework architecture, detailing hyperplane construction and distance-guided generation via VerDisGAN and HorDisGAN.

An auxiliary classifier ($C_{aux}$) is utilized to reconstruct these distances from the generated images, providing a feedback loop that forces the generators to adhere to the specified distance constraints.

Experimental Results and Performance

The authors evaluated DisGAN across four challenging small-scale datasets: Butterfly Mimics (Monarch vs. Viceroy), Asian vs. African Elephants, Breast Ultrasound (Benign vs. Malignant), and COVID-CT scans.

DisGAN consistently outperformed both traditional augmentation (TA) and established GAN variants like ACGAN and CycleGAN. For instance, on the Breast Ultrasound dataset using a ConvNeXt backbone, DisGAN achieved an accuracy of 0.887, a significant improvement over the 0.852 baseline. Similarly, in the elephant classification task, the framework reached an AUC of 0.969, compared to the 0.948 baseline. t-SNE visualizations confirm that DisGAN effectively fills the "gaps" in the feature space, reinforcing the decision boundary and making the classifier more resilient to noise.

Explainability via Class-Difference Maps

Beyond performance gains, DisGAN enhances model interpretability through Class-Difference Maps (CDMs). By calculating the difference between an original image and its mathematical projection onto the hyperplane (the point where $d_v = 0$), the framework can visualize the specific features that drive a classification decision. This provides a transparent look into "why" a model identifies an image as belonging to a particular class, which is vital for high-stakes applications like medical diagnosis.

Conclusion and Future Directions

DisGAN represents a significant step forward in data augmentation for small datasets. By treating the decision boundary as a geometric entity to be filled and refined, it provides a more principled approach to synthetic data generation. While currently optimized for binary classification, future work aims to extend this hyperplane-guided logic to multi-class problems and explore iterative training loops where the classifier and GAN continuously refine one another.

Summary updated

BibTeX
@misc{xiong2023distance,
  author = {Xiangyu Xiong and Yue Sun and Xiaohong Liu and Wei Ke and Chan-Tong Lam and Jiangang Chen and Mingfeng Jiang and Mingwei Wang and Hui Xie and Tong Tong and Qinquan Gao and Hao Chen and Tao Tan},
  title = {{Distance Guided Generative Adversarial Network for Explainable Binary Classifications}},
  year = {2023},
  eprint = {2312.17538},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.17538},
}
Figure