Scholarlink

<p>The performance of classification methods is adversely affected by imbalanced multi-class data. Oversampling is a common solution for addressing imbalanced multi-class classification in data preprocessing. However, excessive oversampling of minority classes results in noise generation and decreases classification accuracy. To solve these problems, a multi-class SMOTEBoost (MSMOTEBoost) method is developed for imbalanced multi-class classification. MSMOTEBoost contains three key components. First, to avoid the generation of noisy examples, the weight of safety (<i>WOS</i>) is designed to select candidate examples for interpolation. Second, to alleviate the excessive interpolation density in certain regions, the weight of multi-class neighbors (<i>WON</i>) is designed to replace the neighborhood random selection process for the dynamic extension of the minority classes. Finally, a multi-class oversampling method fused with AdaBoost.M2 is developed to improve the diversity and robustness of the method. Extensive experiments and statistical analyses of real data are performed to validate the performance of the proposed method. The experimental results demonstrate that MSMOTEBoost is competitive.</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 position oversampling based on ensemble for imbalanced multi-class classification

  • 期刊论文

  • 发表日期：

    2025年4月25日

查看全文
  • Su-Yang Zheng,
  • Chou-Yong Chen,
  • Hong-Jie Li,
  • Chen-Yue Zhang,
  • Zhong-Liang Zhang

摘要

The performance of classification methods is adversely affected by imbalanced multi-class data. Oversampling is a common solution for addressing imbalanced multi-class classification in data preprocessing. However, excessive oversampling of minority classes results in noise generation and decreases classification accuracy. To solve these problems, a multi-class SMOTEBoost (MSMOTEBoost) method is developed for imbalanced multi-class classification. MSMOTEBoost contains three key components. First, to avoid the generation of noisy examples, the weight of safety (WOS) is designed to select candidate examples for interpolation. Second, to alleviate the excessive interpolation density in certain regions, the weight of multi-class neighbors (WON) is designed to replace the neighborhood random selection process for the dynamic extension of the minority classes. Finally, a multi-class oversampling method fused with AdaBoost.M2 is developed to improve the diversity and robustness of the method. Extensive experiments and statistical analyses of real data are performed to validate the performance of the proposed method. The experimental results demonstrate that MSMOTEBoost is competitive.