基于局部集合和差分进化的过抽样方法  

Oversampling Method Based on Local Sets and Differential Evolution

作　　者：罗少甫[1] Evolution

机构地区：[1]重庆航天职业技术学院智能信息学院,重庆400021

出　　处：《统计与决策》2023年第10期23-29,共7页Statistics & Decision

基　　金：重庆市教育委员会科学技术研究项目(KJQN201803001)。

摘　　要：基于噪声过滤的过抽样方法能克服噪声生成,但是他们的噪声侦察技术和合成样本的过程依赖于参数,同时,他们也会过滤掉大量的少数类样本。为了解决该问题,文章提出了一种基于局部集合和差分进化的过抽样方法(OMLSDE)。首先,用该方法计算每个样本的局部集合;其次,用局部集合和不平衡比去发现可疑的噪声;然后,用差分进化去优化可疑的噪声;最后,用局部集合内的随机样本去生成少数类的合成样本。实验证明,OMLSDE优于6个流行的过抽样方法。Noise filtering-based oversampling methods can overcome noise generation,but their noise detection techniques and the process of generating synthetic samples heavily rely on parameters.At the same time,they filter out a large number of minority class samples.In order to overcome the above shortcomings,this paper proposes an oversampling method based on local sets and differential evolution(OMLSDE).Firstly,the local set is calculated for each sample.Secondly,suspicious noise is detected by local sets and the imbalance rate.Then,suspicious noise is optimized by differential evolution.Finally,random samples in local sets are used to create synthetic minority class samples.Experimental results show that OMLSDE outperforms six popular oversampling methods.

关 键 词：不平衡分类 数据预处理 过抽样 局部集合 差分进化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP181[自动化与计算机技术—计算机科学与技术]