一种处理不均衡多分类问题的特征选择集成方法  

An ensemble learning algorithm for feature selection based on solutionto multi-class imbalance data classification

作　　者：宿晨 徐华[1] 崔鑫 王玲娣 SU Chen;XU Hua;CUI Xin;WANG Lingdi(School of Internet of Things Engineering,Jiangnan University,Wuxi 214122,Jiangsu,P.R.China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122

出　　处：《重庆大学学报》2022年第5期125-134,共10页Journal of Chongqing University

基　　金：教育部-新华三集团“云数融合”基金资助项目(2017A13055)。

摘　　要：为解决不均衡多分类问题,提出一种特征选择和AdaBoost的集成方法。首先,数据进行预处理。利用WSPSO算法进行特征选择,根据特征重要性选取初始粒子构建初始种群,使得算法初期就可以沿着正确的搜索方向开展,减少不相关特征的影响。其次,利用AdaBoost算法对于样本权重较敏感的特点,增强对小类样本的关注度。并且利用AUCarea作为评价标准,相对于其他评价标准,AUCarea具有可视化的优点且对较差AUC更加敏感。最后,与其他几种不均衡分类算法在不平衡数据集上进行对比,结果证明该算法可有效处理不均衡多分类问题。In order to solve the problem of unbalanced multi-classification,a feature selection and AdaBoost integration method is proposed.First,the data is preprocessed.The WSPSO algorithm is used to select features,and the initial population is constructed according to the importance of the feature.The initial algorithm can be carried out along the correct search direction to reduce the influence of incoherent features.Secondly,the AdaBoost algorithm is more sensitive to sample weights,and the attention to small samples is enhanced.And using AUCare is used,as the evaluation standard,because compared with other evaluation criteria,AUCare has the advantage of visualization and is more sensitive to poor AUC.Finally,compared with several other unbalanced classification algorithms on the unbalanced data set,the algorithm can effectively deal with the unbalanced multi-classification problem.

关 键 词：不平衡数据 集成学习 ADABOOST 特征选择 多分类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]