计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (2): 176-183.DOI: 10.3778/j.issn.1002-8331.2008-0058

• 模式识别与人工智能 • 上一篇    下一篇

面向不均衡数据的多分类集成算法

崔鑫,徐华,朱亮   

  1. 江南大学 人工智能与计算机学院,江苏 无锡 214122
  • 出版日期:2022-01-15 发布日期:2022-01-18

Multi-classification Ensemble Algorithm for Imbalanced Data

CUI Xin, XU Hua, ZHU Liang   

  1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2022-01-15 Published:2022-01-18

摘要: 为解决不均衡多分类问题,提出了一种基于采样和特征选择的不均衡数据集成分类算法(IDESF)。基分类器的多样性会影响集成算法的分类性能,所以IDESF算法对数据集进行有放回采样+SMOTE的两阶段采样。两阶段采样在保证所得数据集中样本合理性的基础上,增加数据集间的差异性以此隐式地提高基分类器的多样性。两阶段采样同样可以平衡数据分布,防止分类器偏向多数类。在两阶段采样的基础上,IDESF算法引入了数据清洗和特征选择方法,试图进一步提高算法的分类性能。与其他不均衡分类算法在5组不均衡数据集上进行了对比实验,结果表明该算法可以获得较高的AUCarea和G-Mean值,具有较为优异的分类效果。

关键词: 不均衡数据集, 过采样, 数据清洗, 特征选择, 集成算法, 多分类

Abstract: To solve the problem of imbalanced multi-classification, an ensemble classification algorithm(IDESF) based on sampling and feature selection is proposed. The diversity of base classifier affects the classification performance of the ensemble algorithm. Therefore, the IDESF algorithm first performs a two-stage sampling of the data set with replacement sampling+SMOTE. On the basis of ensuring the rationality of the samples in the obtained data set, the two-stage sampling increases the difference among the data sets so as to improve the diversity of the base classifier implicitly. In addition, two-stage sampling can also balance the data distribution and prevent the classifier from favoring the majority classes. Based on two-stage sampling, IDESF algorithm introduces data cleaning and feature selection methods to further improve the classification performance of the algorithm. Finally, the results show that the algorithm can obtain higher AUCarea and G-mean and has better classification effect.

Key words: imbalanced data, over-sampling, data cleaning, feature selection, ensemble algorithm, multi-classification