Scholarlink

<p>Data imbalance is a common issue in many applications, where a significant disparity in sample counts across different classes can cause classifiers to favor the majority class. Traditional oversampling methods often generate samples by directly leveraging the available information of minority classes. In contrast, this study introduces a novel approach, named SDCGAN, which oversamples with a Conditional Generative Adversarial Network (CGAN) where the generator is fed with strengthened distribution information extracted from a curated set of minority samples. This set includes marginal and easily classifiable samples, as well as additional augmented samples derived from these minority samples. The augmentation strategy implemented within the feature space of data preserves and strengthens intrinsic sample distributions and avoids information loss. Extensive experimental results across ten datasets with imbalance ratios (IR) ranging from 1.86 to 23.1 demonstrate that SDCGAN surpasses 13 widely-used resampling techniques in terms of <i>G</i>-mean, <i>F</i>-measure, AUC, and accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GAN-powered oversampling with strengthened sample distribution for class overlapping imbalanced data

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Yulin Zhang,
  • Xiaozhe Wang,
  • Kaiwen Xue,
  • Andrei Zheltkovich,
  • Jionglong Su,
  • Jie Yang

摘要

Data imbalance is a common issue in many applications, where a significant disparity in sample counts across different classes can cause classifiers to favor the majority class. Traditional oversampling methods often generate samples by directly leveraging the available information of minority classes. In contrast, this study introduces a novel approach, named SDCGAN, which oversamples with a Conditional Generative Adversarial Network (CGAN) where the generator is fed with strengthened distribution information extracted from a curated set of minority samples. This set includes marginal and easily classifiable samples, as well as additional augmented samples derived from these minority samples. The augmentation strategy implemented within the feature space of data preserves and strengthens intrinsic sample distributions and avoids information loss. Extensive experimental results across ten datasets with imbalance ratios (IR) ranging from 1.86 to 23.1 demonstrate that SDCGAN surpasses 13 widely-used resampling techniques in terms of G-mean, F-measure, AUC, and accuracy.