基于相对熵的数据流概念漂移检测算法  

A DATA FLOW CONCEPTUAL DRIFT DETECTION ALGORITHM BASED ON RELATIVE ENTROPY

作　　者：杨帆[1] 张永[1,2] 

机构地区：[1]辽宁师范大学计算机与信息技术学院,辽宁大连116081 [2]计算机软件新技术国家重点实验室(南京大学),江苏南京210023

出　　处：《计算机应用与软件》2017年第12期256-259,共4页Computer Applications and Software

基　　金：国家自然科学基金面上项目(61373127)

摘　　要：针对数据流中出现的概念漂移问题,采用决策树作为分类器,提出一种基于相对熵的数据流概念漂移检测算法。提出的算法将分类器的准确率与相对熵作为判断该数据块是否发生概念漂移的标准。通过5个数据集对该方法进行验证,该算法在其中4个数据集上都获得了最优的结果,在另一个数据集上获得了次优结果。实验结果表明采用该方法不仅能够有效地检测概念漂移的发生,而且还能提高分类器的准确率。Aiming at the problem of concept drift in data stream, this paper proposed a conceptual drift detection algorithm based on relative entropy based on decision tree as a classifier. The proposed algorithm combined the accuracy and relative entropy of the classifier as a criterion for judging whether the data block was drilled or not. The method was verified by 5 data sets. The algorithm obtained the optimal result on the four data sets, and the suboptimal result was obtained.on the other data set. The experimental results showed that this method not only detected the occurrence of concept drift effectively, but also improved the accuracy of the classifier.

关 键 词：数据流 概念漂移 相对熵 决策树 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]