一种标记粒化集成的多标记学习算法  

Label-granulated Ensemble Method for Multi-label Learning

作　　者：李峰[1,2] 苗夺谦[1,2] 张志飞[1,2] 罗晟 LI Feng;MIAO Duo-qian;ZHANG Zhi-fei;LUO Sheng(Department of Computer Science and Technology, Tongji University, Shanghai 201804, China;Key Laboratory of Embedded System and Service Computing, Ministry of Education,Tongji University, Shanghai 201804, China)

机构地区：[1]同济大学计算机科学与技术系,上海201804 [2]同济大学嵌入式系统与服务计算教育部重点实验室,上海201804

出　　处：《小型微型计算机系统》2018年第6期1121-1125,共5页Journal of Chinese Computer Systems

基　　金：国家重点研发计划项目(2017YFC0821300)资助;国家自然科学基金项目(61573255;61673301)资助

摘　　要：问题转化型方法和算法适应型方法是多标记学习中主要的两类研究方法,其中问题转化型方法因其独立分类算法得到了广泛的关注,而已有的问题转化型方法存在或忽略标记间相关性,或算法复杂过高,或算法性能不稳定的问题.针对上述不足,基于粒计算的思想,本文提出了一种粒化集成的多标记学习算法.该算法为每个标记划分出一个相关性最大的标记子集,称为关系粒,将标记空间粒化为多个标记子集,该方式考虑到并最大化保留了标记间的相关性,避免了算法复杂度过高,提升了算法性能.随后为每个关系粒训练一个分类模型,最终将各个分类模型的结果集成.实验结果表明相较于对比的三种方法,本文所提算法能取得较好的性能.Problem transformation methods and algorithm adaption methods are two main groups of methods in the multi-label learning.Problem transformation methods have attracted significant attentions,because of its independence on classification algorithms.However,the existing problem transformation methods either ignore the label correlation,or have high computing cost,or unstable predictions.To avoid the above problems,this paper proposed a label-granulated ensemble method for multi-label learning based on granular computing theory. For each label,a label subset consisting of the most related k labels is found,called label-correlated granule,and the label space is granulated into several label subsets,which takes label correlations into consideration as many as possible,avoids high computational complexities and improves the classification performance. A classification for each label-correlated granule is trained,and the prediction achieved by every classification is assembled. The experimental results show that the proposed method can obtain better performance than other three compared methods.

关 键 词：粒计算 集成学习 多标记学习 互信息 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]