互联网 qkzz.net
全刊杂志网:首页 > 女性 > 文章正文
刊社推荐

双层分类策略及其实现算法


□ 凌 萍 周春光

  摘 要:针对常见分类算法在全局和局部区域性能不一致的问题,提出了双层分类策略及其实现算法。双层分类策略的思想是离线地建立全局分类器,当全局分类器决策信用度低于指定阈值时,在线生成局部分类器进行决策修正。实现算法以支持向量机(support vector machine,SVM)和模糊分类器(fuzzy classifier)作为全局与局部分类器,命名为SFC。为全局分类器定义了SVM决策信用度的评估机制,并以此给出局部分类器的启动条件。为局部分类器设计了基于新测度的模糊隶属度函数完成决策修正。实验结果表明,SFC显著提高了单一分类器的性能,可达到较好的分类效果。由此说明双层分类思想是正确且有效的,可作为一种通用思想对应多种具体实现算法。
  关键词:双层分类策略;全局分类器;局部分类器;信用度评估;新测度
  中图分类号:TP139文献标志码:A
  文章编号:1001-3695(2009)06-2074-05
  doi:10.3969/j.issn.1001-3695.2009.06.022
  
  Double-classification strategy and its implementation algorithm
  LING Ping1,2,ZHOU Chun-guang2
  (1.College of Computer Science & Technology, Xuzhou Normal University, Xuzhou Jiangsu 221116, China;2.College of Computer Science & Technology, Jilin University, Changchun 130012, China)
  Abstract:To address the unsteady performance of common classifier in global environment and local region,this paper proposed a double-classification strategy and its implementation algorithm.The idea of the strategy was to create global classifier offline and local classifier online. When the confidence of global classifier’s decision was below some threshold,constructed local classifier to refine the global decision. The implementation algorithm employed support vector machine (SVM) and fuzzy classifier as global and local classifier respectively, named as SFC.Defined the confidence scoring method of SVM decision.Based on that,gave motivation condition of local classifier.For local classifier,defined a new metric to facilitate the work of fuzzy membership function. Empirical evidence shows SFC improves classification performance of the single classifier, and exhibites fine classification ability in applications. SFC’s good behaviors lead to the conclusion that double-classification idea is correct and valid, and it can be realized into multi implementations. ......
很抱歉,暂无全文,若需要阅读全文或喜欢本刊物请联系《计算机应用研究》杂志社购买。
欢迎作者提供全文,请点击编辑
分享:
 

了解更多资讯,请关注“木兰百花园”
分享:
 
精彩图文


关键字
支持中国杂志产业发展,请购买、订阅纸质杂志,欢迎杂志社提供过刊、样刊及电子版。
关于我们 | 网站声明 | 刊社管理 | 网站地图 | 联系方式 | 中图分类法 | RSS 2.0订阅 | IP查询
全刊杂志赏析网 2017