融合粒度划分的回归标签噪声过滤方法  

Granular Ball Label Noise Filtering Method for Regression

作　　者：史颖[1,2] 祁晓博[1] 亓慧 姜高霞[2] 王文剑[3,4] SHI Ying;QI Xiaobo;QI Hui;JIANG Gao-xia;WANG Wen-jian(College of Computer Science and Technology,Taiyuan Normal University,Jinzhong 030619,China;School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;Institute of Intelligent Information Processing,Shanxi University,Taiyuan 030006,China;Department of Network Security,Shanxi Police College,Taiyuan 030401,China)

机构地区：[1]太原师范学院计算机科学与技术学院,山西晋中030619 [2]山西大学计算机与信息技术学院,太原030006 [3]山西大学智能信息处理研究所,太原030006 [4]山西警察学院网络安全保卫系,太原030401

出　　处：《小型微型计算机系统》2024年第8期1832-1838,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(U21A20513,62076154,61906113)资助;山西省专利转化专项计划项目(202302009,202302012)资助;山西省重点研发计划项目(202202020101003)资助;山西省基础研究计划(自由探索类)青年项目(20210302123334)资助.

摘　　要：机器学习算法的性能与数据质量密切相关,然而实际数据中可能存在各类噪声,如何处理标签噪声是机器学习的重要挑战之一.由于回归标签噪声的复杂性,现有噪声过滤方法在处理数值型标签噪声时效果不够理想.本文首先通过标签离散化方式提出一种基于粒度球的数值型标签噪声估计方法,然后给出标签含噪时回归模型的泛化误差估计,并将其作为噪声过滤的目标函数进行优化,最后将所提噪声估计方法与目标函数结合设计了粒度球标签噪声过滤(Granular Ball Label Noise Filtering,GBNF)方法.在标准数据集和真实年龄估计数据集的实验结果表明,该方法识别标签噪声的准确率较高,并取得了比其他噪声过滤比较方法更好的测试性能.The performance of machine learning algorithms is closely related to data quality.However,real datasets may have various types of noise.How to deal with label noise is one of the key challenges in machine learning.Due to the complexity of regression label noise,existing noise filtering methods are not effective enough in dealing with numerical label noise.A granular ball estimate of numerical label noise is developed thorough label discretization.Then the generalization error estimation of regression model on label noise is proposed,and it can be taken as an objective function for noise filtering.Finally,a granular ball label noise filtering(GBNF)method is designed by combining the label noise estimate and the objective function.The experimental results on benchmark datasets and a real age estimation dataset show that the proposed method can accurately recognize the label noise and achieve a lower test error than other noise filters.

关 键 词：回归 标签噪声 粒度球 噪声过滤 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]