不平衡数据分类数据层面方法:现状及研究进展

Data-level methods of imbalanced data classification:status and research development

在不平衡数据的分类中,标准分类器为优化整体的分类误差会牺牲少数类的分类准确率,而实际应用中通常更重视对少数类的准确识别。数据层面方法因其有独立于分类器、泛化能力较强、实现简单等优势,成为解决不平衡数据分类问题的有效策略。围绕不平衡数据分类的数据层面方法开展综述研究,分析造成不平衡数据分类问题的影响因素,从样本空间优化、特征空间优化两个方向对重采样方法及特征选择方法的相关研究进行梳理和评述,并对两类方法进行横向比较。最后提出了需要重点关注的问题和可能的研究机会,以期为不平衡数据分类算法研究及应用提供借鉴和参考。

In the classification of imbalanced data,in order to optimize the overall classification error,the standard classifiers may sacrifice the classification accuracy of the minority class.But more attention be paid to the accurate recognition of the minority class in practical applications.Due to the unique advantages of its high independence from classifiers,strong generalization capability and simplicity,the data-level methods have become more effective strategies to solve the problems of imbalanced data classification.Focusing on the data-level methods of imbalanced data classification,this paper firstly analysed the influencing factors that caused the imbalanced data classification problem.Then it assessed the relevant researches on resampling methods and feature selection methods which corresponding to sample space optimization and feature space optimization respectively,and horizontally compared these two data-level methods.Finally it put forward the issues that need to be focused on and proposed some possible research opportunities,so as to provide references for the algorithm research and applications of im-balanced data classification.