Scholarlink

<p>Class imbalance continues to be a key difficulty in practical classification scenarios, frequently being worsened by the occurrence of minority instances that are noisy and mislabeled. Traditional oversampling techniques like SMOTE, ADASYN, Borderline-SMOTE, and KNNOR produce synthetic samples without specifically considering structural unreliability and label noise, which can result in the propagation of false patterns and reduced classifier accuracy. In this paper, we introduce Graph-based Reliable Adaptive Oversampling (GRASO), a new two-phase framework that solves both class imbalance and minority noise by exploiting the graph structure of minority points. GRASO initially builds a graph of only minority points and calculates a reliability score as a convex combination of label purity and local structural centrality. Unstable nodes with low ranks are discarded, and adaptive sampling is performed only on the most stable node. We tested GRASO on ten standard benchmark datasets with six classifiers under four different amounts of synthetic label noise (0%, 10%, 20%, and 30%). Experimental performance indicates that GRASO performs better than recent popular oversampling techniques with respect to F1 score, G mean, and AUC. Impressively, it has an average F1 score of 0.872 when noise is 0% and is still 0.829 at 30% noise, showing better resistance to noise. Statistical significance tests also further support GRASO’s superiority in all but most settings. These results support the efficacy of graph-based reliability modeling and adaptive sampling in combating the negative impact of class imbalance and noisy samples.</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

Graso: a graph-based reliable oversampling framework for noisy and imbalanced data

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Shivani Goswami,
  • Anil Kumar Singh

摘要

Class imbalance continues to be a key difficulty in practical classification scenarios, frequently being worsened by the occurrence of minority instances that are noisy and mislabeled. Traditional oversampling techniques like SMOTE, ADASYN, Borderline-SMOTE, and KNNOR produce synthetic samples without specifically considering structural unreliability and label noise, which can result in the propagation of false patterns and reduced classifier accuracy. In this paper, we introduce Graph-based Reliable Adaptive Oversampling (GRASO), a new two-phase framework that solves both class imbalance and minority noise by exploiting the graph structure of minority points. GRASO initially builds a graph of only minority points and calculates a reliability score as a convex combination of label purity and local structural centrality. Unstable nodes with low ranks are discarded, and adaptive sampling is performed only on the most stable node. We tested GRASO on ten standard benchmark datasets with six classifiers under four different amounts of synthetic label noise (0%, 10%, 20%, and 30%). Experimental performance indicates that GRASO performs better than recent popular oversampling techniques with respect to F1 score, G mean, and AUC. Impressively, it has an average F1 score of 0.872 when noise is 0% and is still 0.829 at 30% noise, showing better resistance to noise. Statistical significance tests also further support GRASO’s superiority in all but most settings. These results support the efficacy of graph-based reliability modeling and adaptive sampling in combating the negative impact of class imbalance and noisy samples.