《计算机应用》唯一官方网站 ›› 2021, Vol. 41 ›› Issue (12): 3485-3491.DOI: 10.11772/j.issn.1001-9081.2021061026

• 第十八届中国机器学习会议(CCML 2021) • 上一篇    

基于动态概率抽样的标签噪声过滤方法

张增辉1, 姜高霞1,2, 王文剑1,2()   

  1. 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(山西大学),太原 030006
  • 收稿日期:2021-03-16 修回日期:2021-06-29 接受日期:2021-07-21 发布日期:2021-12-28 出版日期:2021-12-10
  • 通讯作者: 王文剑
  • 作者简介:张增辉(1996—),女,山西太原人,硕士研究生,主要研究方向:机器学习
    姜高霞(1987—),男,山西新绛人,副教授,博士,CCF会员,主要研究方向:数据质量分析、机器学习;
  • 基金资助:
    国家自然科学基金资助项目(62076154);山西国际科技合作计划项目(201903D421050);中央引导地方科技发展资金项目(YDZX20201400001224);山西省高等学校科技创新项目(2020L0007)

Label noise filtering method based on dynamic probability sampling

Zenghui ZHANG1, Gaoxia JIANG1,2, Wenjian WANG1,2()   

  1. 1.School of Computer and Information Technology,Shanxi University,Taiyuan Shanxi 030006,China
    2.Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education (Shanxi University),Taiyuan Shanxi 030006,China
  • Received:2021-03-16 Revised:2021-06-29 Accepted:2021-07-21 Online:2021-12-28 Published:2021-12-10
  • Contact: Wenjian WANG
  • About author:ZHANG Zenghui, born in 1996, M. S. candidate. Her research interests include machine learning.
    JIANG Gaoxia, born in 1987, Ph. D., associate professor. His research interests include data quality analysis, machine learning.
  • Supported by:
    the National Natural Science Foundation of China(62076154);the Program of Shanxi Province International Scientific and Technological Cooperation(201903D421050);the Project of Central Government to Guide Local Scientific and Technological Development(YDZX20201400001224);the Scientific and Technological Innovation Program of Higher Education Institutions in Shanxi(2020L0007)

摘要:

在机器学习问题中,数据质量对系统预测的准确性产生了深远的影响。由于信息获取的难度大,人类的认知主观且有限,导致了专家无法准确标记所有样本。而近年来出现的一些概率抽样方法无法避免样本人为划分不合理且主观性较强的问题。针对这一问题,提出一种基于动态概率抽样(DPS)的标签噪声过滤方法,充分考虑各个数据集样本间的差异性,通过统计各个区间内置信度分布频率,分析各个区间内置信度分布信息熵的走势,确定合理阈值。在UCI经典数据集中选取了14个数据集,将所提方法与随机森林(RF)、HARF、MVF、局部概率抽样(LPS)等方法进行了对比实验。实验结果表明,所提出的方法在标签噪声识别和分类泛化上均展示出了较高的能力。

关键词: 标签噪声, 动态概率抽样, 噪声过滤, 标签置信度, 置信度

Abstract:

In machine learning, data quality has a far-reaching impact on the accuracy of system prediction. Due to the difficulty of obtaining information and the subjective and limited cognition of human, experts cannot accurately mark all samples. And some probability sampling methods proposed in resent years fail to avoid the problem of unreasonable and subjective sample division by human. To solve this problem, a label noise filtering method based on Dynamic Probability Sampling (DPS) was proposed, which fully considered the differences between samples of each dataset. By counting the frequency of built-in confidence distribution in each interval and analyzing the trend of information entropy of built-in confidence distribution in each interval, the reasonable threshold was determined. Fourteen datasets were selected from UCI classic datasets, and the proposed algorithm was compared with Random Forest (RF), High Agreement Random Forest Filter (HARF), Majority Vote Filter (MVF) and Local Probability Sampling (LPS) methods. Experimental results show that the proposed method shows high ability on both label noise recognition and classification generalization.

Key words: label noise, Dynamic Probability Sampling (DPS), noise filtering, label confidence, confidence

中图分类号: