Scholarlink

<p>The primary objective of research concerning class imbalance problems revolves around the generation of high-quality data for minority classes. Prior investigations have witnessed various approaches to synthesizing data, resulting in varying data quality. This study introduces a novel oversampling framework, termed the Optimal Oversampling Framework (OOF), which adopts a distinctive perspective. OOF uses optimization algorithms to guide the data generation process, ensuring that new samples are not only similar to the minority class but also exhibit sufficient diversity. Specifically, the method combines initialization and evolutionary strategies to refine the generated samples, while evaluating the similarity of the samples to the minority and majority classes through distance and cosine similarity measures. In addition, OOF prevents premature convergence and ensures that the generated samples maintain uniqueness through diversity judgments and fitness function settings. Finally, OOF selects the best quality samples for oversampling by optimizing the ranking. To demonstrate the effectiveness of OOF, we integrated the Particle Swarm Optimization algorithm with OOF and conducted comparative experiments involving nine different oversampling methods across 21 datasets characterized by high class imbalance ratios. The experimental outcomes validate the success of the OOF approach.</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

Enhancing minority data generation through optimization in imbalanced datasets

  • 期刊论文

  • 发表日期：

    2025年2月10日

查看全文
  • Jiuxiang Song,
  • Chuang Wang,
  • Jizhong Liu

摘要

The primary objective of research concerning class imbalance problems revolves around the generation of high-quality data for minority classes. Prior investigations have witnessed various approaches to synthesizing data, resulting in varying data quality. This study introduces a novel oversampling framework, termed the Optimal Oversampling Framework (OOF), which adopts a distinctive perspective. OOF uses optimization algorithms to guide the data generation process, ensuring that new samples are not only similar to the minority class but also exhibit sufficient diversity. Specifically, the method combines initialization and evolutionary strategies to refine the generated samples, while evaluating the similarity of the samples to the minority and majority classes through distance and cosine similarity measures. In addition, OOF prevents premature convergence and ensures that the generated samples maintain uniqueness through diversity judgments and fitness function settings. Finally, OOF selects the best quality samples for oversampling by optimizing the ranking. To demonstrate the effectiveness of OOF, we integrated the Particle Swarm Optimization algorithm with OOF and conducted comparative experiments involving nine different oversampling methods across 21 datasets characterized by high class imbalance ratios. The experimental outcomes validate the success of the OOF approach.