一种基于原型学习的自适应概念漂移分类方法.PDFVIP

2017年6月 北 京 邮 电 大 学 学 报 Jun.2017 第40卷 第3期 Journal of Beijing University of Posts and Telecommunications Vol.40 No.3 摇 摇 文章编号:1007鄄5321(2017)03鄄0043鄄08 DOI:10.13190/ j.jbupt.2017.03.005 一种基于原型学习的自适应概念漂移分类方法 1 1 1 2 1 苏摇 静 ,摇 裘晓峰 ,摇 李书芳 ,摇 刘道伟 ,摇 张春红 (1郾 北京邮电大学 网络体系构建与融合北京市重点实验室,北京 100876;2郾 中国电力科学研究院,郑州450052) 摘要:为了更准确快速地处理或适应概念漂移,提出了基于原型学习的数据流分类算法,基于发掘并优化现有方法 存在的问题,提出了新的方法模型SyncPrototype,在预测方法、原型判定与更新方法等处理概念漂移问题的关键部 分做出了新的尝试与优化. 实验结果证明,相较于现有方法,SyncPrototype模型在分类性能、概念漂移的响应速度 以及时间性能等方面都有明显提高,能够更加有效处理并适应数据流概念漂移问题. 关摇 键摇 词:数据流;概念漂移;分类 中图分类号:TN992郾53摇 摇 摇 摇 文献标志码:A A Prototype鄄Based Adaptive Concept Drift Classification Method 1 1 1 2 1 SUJing ,摇 QIU Xiao鄄feng ,摇 LI Shu鄄fang ,摇 LIU Dao鄄wei ,摇 ZHANG Chun鄄hong (1郾Beijing Key Laboratory of Network System Architecture and Convergence,Beijing University of Posts and Telecommunications,Beijing 100876,China;2郾China Electric Power Research Institute,Zhengzhou450052,China) Abstract:As a frequent problem that needs to be mainly dealt with in supervised learning scenario of streaming data,theconceptdrift,primarily,occurswhenthedatadistributionorthetargetvariablechan鄄 ges over time. As typical data streams,the research methodwhich real鄄time solves or adaptsto the con鄄 cept drift of data streams can provide strong support for grid security dispatch and stable control of real鄄 time decision鄄making. Foraccurateandquickdealingwithoradaptingtoconceptdrift,aprototype鄄based learning algorithm of data streams classification is discussed. Based on improving the problems which have been explored in existi