Generative Adversarial Minority Oversampling (1903.09730v3)

Abstract: Class imbalance is a long-standing problem relevant to a number of real-world applications of deep learning. Oversampling techniques, which are effective for handling class imbalance in classical learning systems, can not be directly applied to end-to-end deep learning systems. We propose a three-player adversarial game between a convex generator, a multi-class classifier network, and a real/fake discriminator to perform oversampling in deep learning systems. The convex generator generates new samples from the minority classes as convex combinations of existing instances, aiming to fool both the discriminator as well as the classifier into misclassifying the generated samples. Consequently, the artificial samples are generated at critical locations near the peripheries of the classes. This, in turn, adjusts the classifier induced boundaries in a way which is more likely to reduce misclassification from the minority classes. Extensive experiments on multiple class imbalanced image datasets establish the efficacy of our proposal.

• The paper introduces a novel three-player adversarial framework where a convex generator produces synthetic minority samples to challenge both the classifier and the discriminator.
• The method employs a Conditional Transient Mapping Unit and Instance Generation Units to generate samples as convex combinations, preserving the minority class distribution.
• Experimental results on datasets like MNIST and CIFAR10 show improved average class-specific and geometric mean accuracy, demonstrating GAMO's effective handling of imbalanced data.

Generative Adversarial Minority Oversampling

The challenge of class imbalance is prevalent in several real-world applications, where minority class samples are scarce compared to majority class samples, leading to biased model predictions. The paper under discussion introduces the "Generative Adversarial Minority Oversampling" (GAMO) technique to address this issue within deep learning frameworks. It advances prior approaches by proposing an innovative oversampling mechanism specifically tailored for end-to-end deep learning systems, where traditional oversampling methods cannot be directly implemented.

Methodology

GAMO employs a three-player adversarial game comprising a convex generator, a multi-class classifier network, and a real/fake discriminator. The generator is responsible for producing synthetic samples of minority class instances, which are strategically generated as convex combinations of existing samples. The goal is to mislead both the discriminator and the classifier into misclassifying these generated samples. This adversarial interaction ensures that the artificial samples are generated near the edges of class boundaries, facilitating a more precise adjustment of classifier-induced boundaries to mitigate minority class misclassification.

The architecture of the generator is particularly noteworthy. It consists of a Conditional Transient Mapping Unit (cTMU) and Instance Generation Units (IGUs) specific to each minority class. The cTMU maps from a latent space to an intermediate space conditioned on the class label, while each IGU uses this intermediate vector to generate a new sample by means of convex combinations of data points from the minority class.

Significant Contributions

This research distinguishes itself in several ways:

• Adversarial Framework: Unlike typical adversarial classification settings where the generator aligns with the classifier, in GAMO, the generator conspires against both the classifier and discriminator, uniquely tailored to deep learning's end-to-end nature.
• Convex Hull Utilization: By confining the generator's output to the convex hull of real samples, the technique ensures fidelity to the minority class distribution, particularly addressing non-convex class shapes.
• Class-specific Sample Generation: The generator can adaptively oversample various regions within the class distribution based on gradients from the classifier, thereby avoiding excessive parameter tuning typically required in oversampling approaches.

Experimental Findings

The effectiveness of GAMO is demonstrated across multiple imbalanced datasets like MNIST, SVHN, CIFAR10, and SUN397. Using convolutional feature extraction significantly enhanced performance, as evidenced by improved average class-specific accuracy (ACSA) and geometric mean accuracy (GM) scores compared to baseline models and other oversampling techniques like SMOTE and DOS. Furthermore, novel GAMO applications such as GAMO2pix enable realistic image generation from feature space instances, outperforming contemporary models measured by Fréchet Inception Distance (FID).

Implications and Future Directions

GAMO's implications extend both practical and theoretical realms. Practically, it offers an efficient solution to class imbalance in deep learning without extensive parameter tuning. Theoretically, it fosters development in adversarial model training, boundary adjustment techniques, and generative sampled-based learning.

Future work may investigate enhancements in the quality of images generated by GAMO2pix, possibly by exploring more sophisticated architectures or integrating improved GAN variants. Extending GAMO's potential to incorporate new classes dynamically and exploring its applications in reducing GAN boundary distortion present intriguing research directions.

In summary, the GAMO framework encapsulates a robust approach to overcoming class imbalance in deep learning systems, blending adversarial strategies with direct oversampling methods to yield competitive outcomes across diverse imaging datasets. The interplay between the convex generator and adversarial gaming principles exemplifies a significant contribution to deep learning model training in imbalanced settings.