Scholarlink

<p>Class imbalance in datasets is a critical issue that leads to biased and misleading results in machine learning algorithms. To address this, two novel oversampling techniques are proposed: MK-SMOTE (Modified <i>K</i>-Means SMOTE) and M-SMOTE. MK-SMOTE minimizes noise by clustering minority class samples before generating synthetic data, while M-SMOTE balances the dataset through undersampling and iteratively creates synthetic instances based on calculated probabilities, retaining only those instances that are more likely to belong to the minority class. Both techniques were evaluated against <i>K</i>-Means SMOTE using <i>G</i>-Mean, <i>F</i>1-score, and accuracy as evaluation metrics across 22 benchmark datasets. The results show that MK-SMOTE and M-SMOTE consistently outperform <i>K</i>-Means SMOTE. For instance, with the KNN classifier, M-SMOTE achieved an average <i>F</i>1-score, <i>G</i>-Mean, and accuracy of 0.88, 0.89, and 0.89, respectively, compared to 0.87, 0.86, and 0.89 for MK-SMOTE, and 0.84, 0.83, and 0.86 for <i>K</i>-Means SMOTE. When using Logistic Regression, M-SMOTE yielded further improvements, with an average <i>F</i>1-score, <i>G</i>-Mean, and accuracy of 0.90, 0.90, and 0.91, respectively, demonstrating its effectiveness across multiple classifiers. Additionally, M-SMOTE achieved up to a 20% increase in <i>F</i>1-score and <i>G</i>-Mean for certain datasets, highlighting its robustness in addressing class imbalance. These results underscore the potential of the proposed methods to improve model performance and mitigate noise in imbalanced datasets.</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

MK-SMOTE and M-SMOTE: enhanced techniques for handling class imbalance problem

  • 期刊论文

  • 发表日期：

    2025年2月24日

查看全文
  • Asifa Kanwal,
  • Nayyer Masood,
  • Ghulam Mustafa,
  • Maryam Abdul Ghafoor,
  • Samreen Ayaz

摘要

Class imbalance in datasets is a critical issue that leads to biased and misleading results in machine learning algorithms. To address this, two novel oversampling techniques are proposed: MK-SMOTE (Modified K-Means SMOTE) and M-SMOTE. MK-SMOTE minimizes noise by clustering minority class samples before generating synthetic data, while M-SMOTE balances the dataset through undersampling and iteratively creates synthetic instances based on calculated probabilities, retaining only those instances that are more likely to belong to the minority class. Both techniques were evaluated against K-Means SMOTE using G-Mean, F1-score, and accuracy as evaluation metrics across 22 benchmark datasets. The results show that MK-SMOTE and M-SMOTE consistently outperform K-Means SMOTE. For instance, with the KNN classifier, M-SMOTE achieved an average F1-score, G-Mean, and accuracy of 0.88, 0.89, and 0.89, respectively, compared to 0.87, 0.86, and 0.89 for MK-SMOTE, and 0.84, 0.83, and 0.86 for K-Means SMOTE. When using Logistic Regression, M-SMOTE yielded further improvements, with an average F1-score, G-Mean, and accuracy of 0.90, 0.90, and 0.91, respectively, demonstrating its effectiveness across multiple classifiers. Additionally, M-SMOTE achieved up to a 20% increase in F1-score and G-Mean for certain datasets, highlighting its robustness in addressing class imbalance. These results underscore the potential of the proposed methods to improve model performance and mitigate noise in imbalanced datasets.