一种利用SPXY采样的标签噪声主动清洗方法  

Active Label Noise Cleaning Method Based on SPXY Sampling

作　　者：门昌骞[1] 孟晓超 姜高霞[1] 王文剑[1,2] MEN Chang-qian;MENG Xiao-chao;JIANG Gao-xia;WANG Wen-jian(School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;Key Laboratory of Computational Intelligence and Chinese Information Processing,Ministry of Education,Taiyuan 030006,China)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]计算智能与中文信息处理教育部重点实验室,太原030006

出　　处：《小型微型计算机系统》2021年第9期1865-1870,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62076154,U1805263)资助;山西省国际合作重点研发计划项目(201903D421050)资助;山西省自然科学基金项目(201901D111030)资助。

摘　　要：基于主动学习的标签噪声清洗方法(Active label noise cleaning,ALNC)是一种通过主动学习筛选疑似噪声样本,进而交给人工专家进行再标记的标签噪声清洗方法.虽然该方法既有很好的噪声识别效果又能保持原有数据的完整性,但仍存在人工额外标记代价较高的问题,即筛选出的疑似噪声样本中存在一定比例的正常样本.为了解决这一问题,降低标签噪声清洗过程中的人工额外检验代价,本文提出了一种基于SPXY(Sample Set Partitioning based on Joint X-Y Distance Sampling)采样的标签噪声主动清洗方法(Active label noise cleaning based on SPXY,SPXYALNC),该方法在主动学习筛选疑似噪声样本的过程中结合了SPXY采样方法,这样既考虑了样本的不确定性,又考虑了样本的代表性,并且在原有标准数据集上针对分类问题进行了实验,实验结果表明该方法在保持原有噪声识别效果的同时可以明显降低人工额外检验代价.Active label noise cleaning(ALNC)is a label noise cleaning method which select the suspected noise samples through active learning and gives them to the artificial experts for re-labeling.Although this method has good noise recognition effect and can maintain the integrity of the original data,it still exists the problem of high additional manual labeling cost,that is,there may be a certain proportion of nominal samples in the selected suspected noise samples.In order to solve this problem,reducing the additional expert examining cost in the process of label noise cleaning,this paper proposes an active label noise cleaning method based on SPXY sampling(SPXYALNC)Jthis method combines SPXY sampling in the process of selecting suspected noise samples in active learning from this,the uncertainty and representativeness of samples are both considered,and having experiments in the same standard data sets for classification problems Jthe experiment results show that this method can significantly reduce artificial additional examining cost while maintaining the original noise recognition effect.

关 键 词：标签噪声 噪声清洗 主动学习 SPXY采样 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]