计算机与现代化

• 网络与通信 • 上一篇    下一篇

一种面向反馈网络的因果特征选择算法及其应用

  

  1. (广东电网有限责任公司信息中心,广东广州510000)
  • 收稿日期:2019-05-07 出版日期:2019-12-11 发布日期:2019-12-11
  • 作者简介:彦逸(1987-),男,湖北安陆人,工程师,硕士,研究方向:因果关系,机器学习,信息通信,E-mail: yanyi@gdxx.csg.cn; 李波(1989-),男,工程师,硕士,研究方向:机器学习,调度监控,E-mail: libo@gdxx.csg.cn; 陈守明(1986-),男,高级工程师,硕士,研究方向:机器学习,信息安全,E-mail: chenshouming@gdxx.csg.cn; 林强(1977-),男,高级工程师,硕士,研究方〖JP2〗向:信息安全,E-mail: linqiang@gdxx.csg.cn; 黄巨涛(1970-),男,高级工程师,硕士,研究方向:软件工程,E-mail: huangjutao@gdxx.csg.cn; 温柏坚(1963-),男,教授级高级工程师,博士,研究方向:信息通信,软件工程,E-mail: wenbojian@gdxx.csg.cn。
  • 基金资助:
    国家自然科学基金资助项目(61876043); 广东电网有限责任公司信息中心项目(037800KK52170002)

A Causal Feature Selection Algorithm for Feedback Networks and Its Applications

  1. (Information Center, Guangdong Power Grid, Guangzhou 510000, China)
  • Received:2019-05-07 Online:2019-12-11 Published:2019-12-11

摘要: 针对2种主流的基于马尔科夫毯(Markov Blanket)和基于信息理论(Information-theoretic)的特征选择策略无法有效解决具有反馈机制的多层网络下的问题,提出一种面向反馈多层网络的因果特征选择方法。该方法首先利用D-separation准则找到目标节点T的邻居节点,即邻居特征Ne(T),然后对目标节点与其余特征求互信息,找出互信息靠前的且不被集合Ne(T)中元素D-separation的特征集合R,最后合并Ne(T)和R即为目标节点对应的特征。该方法有效地避免了基于马尔科夫毯的在反馈网络下特征选择错误和多层网络下最大互信息的特征选择错误的问题。与2种经典的策略在大型电力营销系统中典型告警预测进行对比,相较于主流的特征选择方案,实验结果均表明该方法对于电力营销系统的预测告警特征选择更加有效。

关键词: 特征选择, 因果网络, 马尔科夫毯, D-分离, 互信息, 电力营销系统

Abstract: As the two feature selection strategies of Markov blanket based methods and information-theoretic based methods often fail to solve the feature selection problem under the multi-layer network with feedback mechanism, a causal feature selection method for feedback multi-layer networks is proposed. The method first uses the D-separation method to find the neighbor node of the target node T, that is, the neighbor feature Ne(T). Then, the mutual information of the target node and the remaining features is obtained, and the feature set R of the element D-separation of the mutual information and not the set Ne(T) is found, and finally the Ne(T) and R are merged as the target node. The method effectively avoids the problem of feature selection error based on Markov blanket under feedback network and feature selection of maximum mutual information under multi-layer network. Compared with the two classic methods in the typical warning of power marketing system, the experimental results show that the method is more effective.

Key words: feature selection, causal network, Markov blanket, D-separation, mutual information, electric power marketing system

中图分类号: