Scholarlink

Class Imbalance Problem is one of the prevalent issues where certain classes have fewer instances than others. While much research has focused on the binary class imbalance problem, real-world datasets often contain multiple imbalanced classes. This poses challenges for model training, performance evaluation, and generalization. This paper focuses on handling the multi-class imbalance problem by using the over-sampling technique Borderline Synthetic Minority Over-Sampling Technique (BLSMOTE) with three ensemble-based techniques: Stacking, Bagging and Boosting. Stacking ensemble is designed using Support Vector Machine (SVM) and Random Forests (RF) at base level and Logistic Regression (LR) at meta level. Proposed approaches- BLSMOTE-Stacking, BLSMOTE-Bagging and BLSMOTE-Boosting are evaluated on seven multiclass imbalanced datasets taken from KEEL repository using three key metrics: MAUC (Multiclass-Area under Curve), F1-score and AUPR (Area under Precision Recall curve). It has been noted that BLSMOTE-Stacking excels the other methods with the highest value obtained for MAUC at 99.9%.

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

Addressing Multiclass Imbalance Problem: A Synergistic Approach Using BLSMOTE and Ensemble Method

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kashika Wadhwa,
  • Ritika Kumari,
  • Anjana Gosain

摘要

Class Imbalance Problem is one of the prevalent issues where certain classes have fewer instances than others. While much research has focused on the binary class imbalance problem, real-world datasets often contain multiple imbalanced classes. This poses challenges for model training, performance evaluation, and generalization. This paper focuses on handling the multi-class imbalance problem by using the over-sampling technique Borderline Synthetic Minority Over-Sampling Technique (BLSMOTE) with three ensemble-based techniques: Stacking, Bagging and Boosting. Stacking ensemble is designed using Support Vector Machine (SVM) and Random Forests (RF) at base level and Logistic Regression (LR) at meta level. Proposed approaches- BLSMOTE-Stacking, BLSMOTE-Bagging and BLSMOTE-Boosting are evaluated on seven multiclass imbalanced datasets taken from KEEL repository using three key metrics: MAUC (Multiclass-Area under Curve), F1-score and AUPR (Area under Precision Recall curve). It has been noted that BLSMOTE-Stacking excels the other methods with the highest value obtained for MAUC at 99.9%.