基于可信多数投票的快速概念漂移检测  

Fast Detection of Concept Drifts Based on Confident Majority Voting

作　　者：文益民[1,2] 王耀南[1] 张莹[3] 

机构地区：[1]湖南大学电气与信息工程学院,湖南长沙410082 [2]湖南工业职业技术学院,湖南长沙410208 [3]湘潭大学信息工程学院,湖南湘潭411105

出　　处：《湖南大学学报（自然科学版）》2010年第6期36-40,共5页Journal of Hunan University:Natural Sciences

基　　金：国家863项目资助(2007AA04Z244);国家自然科学基金重点资助项目(60835004);湖南省教育厅科学研究项目(09C1134)

摘　　要：数据流因具有数据持续到达,概念漂移产生时刻无法预测、概念的数量不确定等特征,使得滑动窗口的大小很难事先确定,滑动窗口包含概念的数量对概念漂移检测存在影响.本文提出了基于可信多数投票的快速概念漂移检测算法(CMV_SEA),该算法使用SEA算法中的基分类器淘汰方法,使用可信多数投票实现滑动窗口中基分类器的集成.仿真实验表明:相比于SEA算法,CMV_SEA算法提高了泛化能力;能在新概念产生的第一时间内检测到概念漂移;对概念漂移的检测能力和新概念的学习能力不受滑动窗口大小的影响.Data stream is characterized by the continuous arrival of data,the unpredictability of the time of concept drifts and the uncertain number of concepts,making it difficult to predict the size of the window in advance,but the number of concepts in the window affects the detection of concept drifts.This paper proposed an algorithm for the fast detection of concept drifts in data streams by taking a confident majority voting strategy（CMV_SEA）.This algorithm replaced base classifiers in a window like SEA and used majority voting strategy to ensemble all base classifiers in the window.The experiment results illustrate that CMV_SEA can promote prediction accuracy and detect concept drifts as soon as a new concept comes up,and its ability to detect and learn a new concept is not influenced by the size of the window.

关 键 词：学习系统 数据流 概念漂移 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]