Li Ying, Yang Yali, Song Peihua, Duan Lian, Ren Rui
School of Logistics Management and Engineering, Nanning Normal University, Nanning, 530001, Guangxi, China.
Guangxi Colleges and Universities Key laboratory of Intelligent Logistics Technology, Nanning Normal University, Nanning, 530001, Guangxi, China.
Sci Rep. 2025 Jul 2;15(1):23521. doi: 10.1038/s41598-025-09506-w.
Class imbalance in datasets often degrades the performance of classification models. Although the Synthetic Minority Over-sampling Technique (SMOTE) and its variants alleviate this issue by generating synthetic samples, they frequently overlook local density and distribution characteristics. Consequently, developing methods that incorporate local spatial information to synthesize samples that better preserve the original data distribution is critical for improving model robustness in class-imbalanced scenarios. To address this gap, we propose an enhanced SMOTE algorithm (ISMOTE), which modifies the spatial constraints for synthetic sample generation. Unlike SMOTE, the proposed method first generates a base sample between two original samples. Then the Euclidean distance between the two samples is multiplied by a random number to generate a random quantity. This random quantity is added or subtracted based on the distance between the base sample and the original samples, ensuring that new samples are generated around the two original samples. By adaptively expanding the synthetic sample generation space, ISMOTE effectively alleviates distortions in local data distribution and density. This study compared the ISMOTE algorithm with seven mainstream oversampling algorithms, using three classifiers on thirteen public datasets from the KEEL, UCI, and Kaggle databases. Comparative analysis of 2D and 3D scatter plots revealed that ISMOTE yields more realistic data distributions. Experimental results demonstrated relative improvements in classifier performance, with F1-score, G-mean, and AUC increasing by 13.07%, 16.55%, and 7.94%, respectively. Furthermore, ISMOTE's parameter adaptability enables its application to multi-class imbalanced datasets.
数据集中的类别不平衡问题常常会降低分类模型的性能。尽管合成少数类过采样技术（SMOTE）及其变体通过生成合成样本缓解了这一问题，但它们常常忽略局部密度和分布特征。因此，开发能够纳入局部空间信息以合成更好保留原始数据分布的样本的方法，对于提高类别不平衡场景下模型的鲁棒性至关重要。为了弥补这一差距，我们提出了一种增强型SMOTE算法（ISMOTE），它修改了合成样本生成的空间约束。与SMOTE不同，该方法首先在两个原始样本之间生成一个基础样本。然后将两个样本之间的欧几里得距离乘以一个随机数，生成一个随机量。根据基础样本与原始样本之间的距离对这个随机量进行加减，确保在两个原始样本周围生成新样本。通过自适应地扩展合成样本生成空间，ISMOTE有效地减轻了局部数据分布和密度的扭曲。本研究将ISMOTE算法与七种主流过采样算法进行了比较，在来自KEEL、UCI和Kaggle数据库的13个公共数据集上使用了三种分类器。对二维和三维散点图的对比分析表明，ISMOTE产生的数据分布更加真实。实验结果表明分类器性能有相对提升，F1分数、G均值和AUC分别提高了13.07%、16.55%和7.94%。此外，ISMOTE的参数适应性使其能够应用于多类别不平衡数据集。
