Scholarlink

<p>The paper explores the class imbalance problem in classification models which show bias towards the majority class and perform poorly for the minority class. The proposed approach combines Rough Set Theory (RST) and Evolutionary Algorithm (EA) to create an innovative resampling framework. Undersampling is achieved by removing ambiguous majority class samples from the boundary region of the rough set. In oversampling, synthetic minority samples are generated in the positive region applying EAs guided by Kullback-Leibler (KL) divergence and Attribute Dependency (AD) based two objective functions. RST improves the class boundary clarity by filtering out ambiguous majority samples. In contrast, AD ensures that the generated synthetic samples retain feature similarity and align closely with the distribution of minority samples in the positive region. KL divergence further refines the process by reducing the distributional gap between the original and synthetic minority samples. Experiments on well-known datasets indicate that the novel framework improves class balance and classification performance by effectively reducing bias towards the majority class. As a result, we get better generalization and improve minority class recognition. The proposed method has been validated with the comparative analysis of existing resampling techniques, which invariably gives better performance in accuracy, F-score and other measures, albeit with a slight increase in time complexity. The method has potential applications in various real-world classification tasks.</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

Rough set theory and multi objective evolutionary algorithm based undersampling and oversampling framework towards class imbalance problem

  • 期刊论文

  • 发表日期：

    2025年8月13日

查看全文
  • Mehwish Naushin,
  • Asit Kumar Das

摘要

The paper explores the class imbalance problem in classification models which show bias towards the majority class and perform poorly for the minority class. The proposed approach combines Rough Set Theory (RST) and Evolutionary Algorithm (EA) to create an innovative resampling framework. Undersampling is achieved by removing ambiguous majority class samples from the boundary region of the rough set. In oversampling, synthetic minority samples are generated in the positive region applying EAs guided by Kullback-Leibler (KL) divergence and Attribute Dependency (AD) based two objective functions. RST improves the class boundary clarity by filtering out ambiguous majority samples. In contrast, AD ensures that the generated synthetic samples retain feature similarity and align closely with the distribution of minority samples in the positive region. KL divergence further refines the process by reducing the distributional gap between the original and synthetic minority samples. Experiments on well-known datasets indicate that the novel framework improves class balance and classification performance by effectively reducing bias towards the majority class. As a result, we get better generalization and improve minority class recognition. The proposed method has been validated with the comparative analysis of existing resampling techniques, which invariably gives better performance in accuracy, F-score and other measures, albeit with a slight increase in time complexity. The method has potential applications in various real-world classification tasks.