关键词: 不平衡数据, CFSFDP聚类算法, 合成过采样, 边界样本

Abstract: As a major challenge in the classification task, the class imbalance problem stems from the significant imbalance between the number of majority and minority samples in the training dataset. This imbalance not only affects the generalization ability of the classifier, but also may lead to a significant decrease in the recognition accuracy of minority samples. Oversampling techniques, especially SMOTE and its variants, serve as an effective means of mitigating such problems by generating additional minority samples to balance the dataset. However, these methods have limitations such as the potential introduction of noise in sample generation, insufficient sample diversity, and insufficient attention to boundary regions. In view of the key role of boundary samples in classification decision-making and their susceptibility to classifier misjudgment, this paper proposes an innovative oversampling strategy to accurately identify boundary samples and generate high-quality new samples around them. Firstly, the CFSFDP clustering algorithm is used to calculate the local density of each minority sample by virtue of its ability to identify the local density peak, and then the samples located at the classification boundary are screened out. Subsequently, by calculating the Euclidean distance between these boundary samples and their nearest majority class samples, a suitable spherical region is defined for each boundary sample, which not only covers the potential distribution range of the boundary samples, but also avoids excessive overlap with the majority class samples. After determining the boundary sample and its corresponding spherical region, a new synthetic sample is randomly generated in the region. This step not only increases the diversity of minority samples, but also makes the generated samples closer to the real boundary distribution, which helps the classifier to better learn the complex features of minority classes. To verify the effectiveness of the proposed method, we comprehensively compare it with nine existing oversampling methods on 32 real-world imbalance datasets. Experimental results show that the proposed method performs well in multiple evaluation indexes.

Key words: unbalanced data, CFSFDP clustering algorithm, synthetic oversampling, boundary samples