Scholarlink

<p>The combination of class imbalance and class overlap represents one of the most critical challenges in machine learning. While early studies identified class overlap as a key factor that aggravates classification challenges in imbalanced domains, subsequent research has largely examined class imbalance and class overlap separately. As a result, their combined effect in multi-class scenarios remains insufficiently explored. Moreover, the lack of a unified investigation into overlap-mitigating data-level approaches limits their systematic evaluation. To address this gap, this paper presents a unified and structured overview of imbalanced data classification from the perspective of class overlap, with a particular focus on multi-class scenarios. We begin by revisiting the fundamental definitions and implications of both phenomena and highlighting their joint impact on classifier performance. We then provide a structured synthesis of data-level preprocessing methods, particularly sampling strategies (oversampling, undersampling, and hybrid sampling), as well as feature selection approaches, to elucidate their respective roles in mitigating challenges caused by class overlap. Finally, we discuss open challenges and outline promising future directions for multi-class imbalanced learning under overlapping class distributions.</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

Class overlap in imbalanced learning: A data-level perspective and comprehensive review

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月28日

查看全文
  • Yujiang Wang,
  • Marshima Mohd Rosli,
  • Norzilah Musa

摘要

The combination of class imbalance and class overlap represents one of the most critical challenges in machine learning. While early studies identified class overlap as a key factor that aggravates classification challenges in imbalanced domains, subsequent research has largely examined class imbalance and class overlap separately. As a result, their combined effect in multi-class scenarios remains insufficiently explored. Moreover, the lack of a unified investigation into overlap-mitigating data-level approaches limits their systematic evaluation. To address this gap, this paper presents a unified and structured overview of imbalanced data classification from the perspective of class overlap, with a particular focus on multi-class scenarios. We begin by revisiting the fundamental definitions and implications of both phenomena and highlighting their joint impact on classifier performance. We then provide a structured synthesis of data-level preprocessing methods, particularly sampling strategies (oversampling, undersampling, and hybrid sampling), as well as feature selection approaches, to elucidate their respective roles in mitigating challenges caused by class overlap. Finally, we discuss open challenges and outline promising future directions for multi-class imbalanced learning under overlapping class distributions.