Scholarlink

<p>Imbalanced datasets pose a significant challenge for existing <i>k</i>-nearest neighbor (<i>k</i>NN)-based methods, which often exhibit a bias toward the majority class and therefore struggle to maintain classification performance on minority-class instances. Although several strategies have been proposed to address this issue, these approaches often fail to effectively emphasize informative minority class instances in the neighborhood. To address this limitation, we propose Proximity-weighted Evidential Dynamic Neighbors (PwEDN), a method that dynamically accumulates the neighbors of a query instance by prioritizing minority-class samples while taking the underlying class distribution into account. In PwEDN, each neighbor is treated as a piece of evidence supporting the classification decision. PwEDN introduces a belief function to quantify the strength of each piece of evidence. It then applies Dempster-Shafer theory to aggregate the evidence weights and classifies the query instance to the class with maximum likelihood. Besides these, we investigate the impact of distance measures on the performance of <i>k</i>NN variants. Extensive experiments over thirty-two (32) benchmark imbalanced datasets demonstrate that PwEDN outperforms nine (9) familiar and state-of-the-art methods, including data-oriented techniques and <i>k</i>NN variants. Furthermore, the Wilcoxon signed-rank test indicates the statistical superiority of PwEDN over all competitive methods in terms of AUPRC at the 5% level of significance.</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

An Evidential Dynamic Neighborhood Approach to Learn from Imbalanced Data

  • 期刊论文

  • 发表日期：

    2026年6月20日

查看全文
  • Md. Eusha Kadir,
  • Md. Mumtahin Habib Ullah Mazumder,
  • Muhammad Mahbub Alam,
  • Mohammad Shoyaib

摘要

Imbalanced datasets pose a significant challenge for existing k-nearest neighbor (kNN)-based methods, which often exhibit a bias toward the majority class and therefore struggle to maintain classification performance on minority-class instances. Although several strategies have been proposed to address this issue, these approaches often fail to effectively emphasize informative minority class instances in the neighborhood. To address this limitation, we propose Proximity-weighted Evidential Dynamic Neighbors (PwEDN), a method that dynamically accumulates the neighbors of a query instance by prioritizing minority-class samples while taking the underlying class distribution into account. In PwEDN, each neighbor is treated as a piece of evidence supporting the classification decision. PwEDN introduces a belief function to quantify the strength of each piece of evidence. It then applies Dempster-Shafer theory to aggregate the evidence weights and classifies the query instance to the class with maximum likelihood. Besides these, we investigate the impact of distance measures on the performance of kNN variants. Extensive experiments over thirty-two (32) benchmark imbalanced datasets demonstrate that PwEDN outperforms nine (9) familiar and state-of-the-art methods, including data-oriented techniques and kNN variants. Furthermore, the Wilcoxon signed-rank test indicates the statistical superiority of PwEDN over all competitive methods in terms of AUPRC at the 5% level of significance.