Scholarlink

<p>Rapidly advancing hardware technology in recent years is significantly accelerating widespread application of deep learning and machine learning algorithms across biomedical and oncological research domains. Although such algorithms have demonstrated superior performance in those domains, their effectiveness in actual biomedical applications is severely constrained by class imbalance issues. To address this problem, this study proposes a novel method termed BSMOTE-DCGAN-EL, which integrates a dual-channel autoencoder conditional generative adversarial network (DCGAN) structure with Borderline Synthetic Minority Oversampling Technique (BSMOTE) generating diverse synthetic minority class samples. Specifically, the proposed DCGAN model leverages manifold features extracting from the original data and their corresponding membership functions (MF), encapsulating rich latent information as conditional features. Furthermore, a bagging-based ensemble learning (EL) algorithm is deployed for mitigating biases toward majority-class samples introduced by traditional classification methods. To validate the effectiveness of the proposed BSMOTE-DCGAN-EL algorithm, experimentation was conducted employing twelve real-world imbalanced biomedical datasets across one widely adopted predictive model. Experimental results demonstrate the proposed method significantly outperforms six state-of-the-art oversampling algorithms and two ablation studies across four evaluation metrics: G-mean, F1, IBA, and AUC. At imbalance ratios of 5, 10, 15, and 20, the proposed approach achieved average improvements of 26.273%, 25.313%, 23.877%, and 21.906%, respectively. Furthermore, statistical analyses employing Friedman and Nemenyi post-hoc tests confirmed the significant superiority of the proposed BSMOTE-DCGAN-EL algorithm over the other six oversampling methods in highly imbalanced data scenarios.</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

Manifold-guided SMOTified dual-channel conditional GAN improving highly-imbalanced biomedical data classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月22日

查看全文
  • Liang-Sian Lin,
  • Chien-Hsin Lin,
  • Hsin-Yu Chang,
  • Hui-Chi Chuang,
  • Chih-Ching Liu

摘要

Rapidly advancing hardware technology in recent years is significantly accelerating widespread application of deep learning and machine learning algorithms across biomedical and oncological research domains. Although such algorithms have demonstrated superior performance in those domains, their effectiveness in actual biomedical applications is severely constrained by class imbalance issues. To address this problem, this study proposes a novel method termed BSMOTE-DCGAN-EL, which integrates a dual-channel autoencoder conditional generative adversarial network (DCGAN) structure with Borderline Synthetic Minority Oversampling Technique (BSMOTE) generating diverse synthetic minority class samples. Specifically, the proposed DCGAN model leverages manifold features extracting from the original data and their corresponding membership functions (MF), encapsulating rich latent information as conditional features. Furthermore, a bagging-based ensemble learning (EL) algorithm is deployed for mitigating biases toward majority-class samples introduced by traditional classification methods. To validate the effectiveness of the proposed BSMOTE-DCGAN-EL algorithm, experimentation was conducted employing twelve real-world imbalanced biomedical datasets across one widely adopted predictive model. Experimental results demonstrate the proposed method significantly outperforms six state-of-the-art oversampling algorithms and two ablation studies across four evaluation metrics: G-mean, F1, IBA, and AUC. At imbalance ratios of 5, 10, 15, and 20, the proposed approach achieved average improvements of 26.273%, 25.313%, 23.877%, and 21.906%, respectively. Furthermore, statistical analyses employing Friedman and Nemenyi post-hoc tests confirmed the significant superiority of the proposed BSMOTE-DCGAN-EL algorithm over the other six oversampling methods in highly imbalanced data scenarios.