Scholarlink

<p>In recent decades, the class imbalanced problem has remained a prevalent challenge in machine learning and data mining. While SMOTE and its variants serve as effective solutions, they exhibit notable disadvantages, including insufficient diversity and an overly concentrated distribution of synthetic samples. To address these problems, we propose a novel oversampling method called Weighted Triangular Regions-Based Oversampling Technique(WTO). WTO synthesizes sample in weighted triangular regions, which not only enhances the diversity of the generated samples but also effectively ensures uniform density distribution. Additionally, we take full advantage of triangular regions within majority samples to generate new samples, which further enhance the classification performance. We conducted experiments on 15 real-world datasets against 11 oversampling methods and maintained superior performance across multiple classifiers. The experimental results demonstrate that our WTO can effectively mitigate the class imbalanced problem and enhance classifier performance. Code is available at <a href="https://github.com/Echooo523/WTO.">https://github.com/Echooo523/WTO.</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

WTO: weighted triangular regions-based oversampling technique for imbalanced classification

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Jianjian Yan,
  • Qicheng Qian,
  • Zhaoqi Yuan,
  • Ziyi Wang,
  • Min Deng,
  • Benting Wan

摘要

In recent decades, the class imbalanced problem has remained a prevalent challenge in machine learning and data mining. While SMOTE and its variants serve as effective solutions, they exhibit notable disadvantages, including insufficient diversity and an overly concentrated distribution of synthetic samples. To address these problems, we propose a novel oversampling method called Weighted Triangular Regions-Based Oversampling Technique(WTO). WTO synthesizes sample in weighted triangular regions, which not only enhances the diversity of the generated samples but also effectively ensures uniform density distribution. Additionally, we take full advantage of triangular regions within majority samples to generate new samples, which further enhance the classification performance. We conducted experiments on 15 real-world datasets against 11 oversampling methods and maintained superior performance across multiple classifiers. The experimental results demonstrate that our WTO can effectively mitigate the class imbalanced problem and enhance classifier performance. Code is available at https://github.com/Echooo523/WTO.