Scholarlink

<p>A novel loss function based on stepwise gradient penalty has been proposed to address model bias in multi-class imbalanced data classification. The method integrates a power-exponential function as a penalty factor into the cross-entropy loss, and matches the corresponding gradient penalty according to the frequency of labels for each class. We analyze the rationality of the method from the perspective of gradients and demonstrate that it is the universal framework of the current mainstream imbalanced data classification loss functions. Finally, various types of multi-class imbalanced datasets, including linearly imbalanced datasets, stepped imbalanced datasets, and datasets with consistent/inconsistent training and test distributions, are constructed for classification experiments on the SVHN, CIFAR-10, and Caltech-101 datasets, respectively. The results on all metrics demonstrate the competitiveness of this method. Our code is available at <a href="https://github.com/gzmtzly/SGPL.">https://github.com/gzmtzly/SGPL.</a></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

A Novel Loss Based on Stepwise Gradient Penalty for Multi-Class Imbalanced Data Classification

  • 期刊论文

  • 发表日期：

    2025年7月25日

查看全文
  • Bimo Ren,
  • Jingyi Feng,
  • Linyong Zhou

摘要

A novel loss function based on stepwise gradient penalty has been proposed to address model bias in multi-class imbalanced data classification. The method integrates a power-exponential function as a penalty factor into the cross-entropy loss, and matches the corresponding gradient penalty according to the frequency of labels for each class. We analyze the rationality of the method from the perspective of gradients and demonstrate that it is the universal framework of the current mainstream imbalanced data classification loss functions. Finally, various types of multi-class imbalanced datasets, including linearly imbalanced datasets, stepped imbalanced datasets, and datasets with consistent/inconsistent training and test distributions, are constructed for classification experiments on the SVHN, CIFAR-10, and Caltech-101 datasets, respectively. The results on all metrics demonstrate the competitiveness of this method. Our code is available at https://github.com/gzmtzly/SGPL.