电子学报 ›› 2022, Vol. 50 ›› Issue (3): 585-597.DOI: 10.12263/DZXB.20210094

• 学术论文 • 上一篇    下一篇

面向概念漂移和类不平衡数据流的在线分类算法

陆克中, 陈超凡, 蔡桓, 吴定明   

  1. 深圳大学计算机与软件学院, 广东 深圳 518060
  • 收稿日期:2021-01-12 修回日期:2021-12-14 出版日期:2022-03-25 发布日期:2022-03-25
  • 作者简介:陆克中 男,1982年生.博士,教授.主要研究方向为大数据计算、并行计算和数据挖掘.
    陈超凡 男,1998年生.硕士研究生在读.主要研究方向为机器学习,深度学习以及数据分析.
    蔡 桓 男,1993年生.硕士研究生.主要研究方向为数据挖掘和机器学习.
    吴定明(通讯作者) 女,1982年生.博士, 副教授.主要研究方向为大数据计算和数据挖掘.E-mail1:dingming@szu.edu.cn
  • 基金资助:
    国家自然科学基金项目(61502310);广东省自然科学基金(2019A1515011721);深圳市基础研究资助项目(20200806102941001)

Online Classification Algorithm for Concept Drift and Class Imbalance Data Stream

LU Ke-zhong, CHEN Chao-fan, CAI Huan, WU Ding-ming   

  1. College of Computer Science and Software Engineering of Shenzhen University,Shenzhen,Guangdong 518060,China
  • Received:2021-01-12 Revised:2021-12-14 Online:2022-03-25 Published:2022-03-25

摘要:

数据流是大数据的重要形式,数据流分类是数据挖掘的重要任务之一,该任务在现实生活中有着巨大的应用前景,因此得到了研究者们的广泛关注.概念漂移和类不平衡是影响数据流分类性能的两个核心问题,但目前大多数算法都只考虑处理两者之一,并且大多数算法过于理想,只能在人工设置的数据流上才能发挥较好的性能,无法适用于复杂的真实数据流.针对这一问题,提出了一种同时处理概念漂移和类不平衡复杂数据流的算法——具有自适应遗忘因子的加权在线顺序极限学习机集成算法.该算法首先融合加权机制和遗忘机制,初步提出具有遗忘机制的加权在线顺序极限学习机算法.为了更好地适应复杂数据流,进一步以初步算法为基分类器,设计包含自适应遗忘因子和概念漂移检测机制的在线集成策略.大量仿真实验表明,所提算法在所有数据集上都取得了最佳的Gmean值,具有更好的概念漂移和类不平衡适应能力,表现出了更稳定、更平衡以及更准确的分类效果.

关键词: 数据流分类, 概念漂移, 类不平衡, 在线学习, 极限学习机

Abstract:

Data stream is an important form of big data, and data stream classification is one of the most important tasks in data mining. This task finds wide application in our life, so it has been attracting great attention of researchers. Concept drift and class imbalance are two main issues that affect the performance of data stream classification algorithms. However, most solutions only address one of these two issues. Even worse, most algorithms can only achieve good performance on data streams under manual settings and cannot be applied to real complex data streams. To solve this problem, an ensemble algorithm of weighted online sequential extreme learning machine with adaptive forgetting factor is proposed to deal with both conceptual drift and imbalance on complex data streams. The proposed algorithm is a weighted online sequential limit learning machine that integrates a weighting mechanism and a forgetting mechanism. In order to adapt to complex data streams, an online integration strategy including adaptive forgetting factor and concept drift detection mechanism was designed as a classifier. Extensive simulation experiments show that the proposed algorithm achieves the best Gmean value on all data sets, has the ability to deal with concept drift and class imbalance, and presents stable, balanced and accurate classification effects.

Key words: data stream classification, concept drift, class imbalance, online learning, extreme learning machine

中图分类号: