结合ReliefF和互信息的多标签特征选择算法  

Multi-label Feature Selection Algorithm Based on ReliefF and Mutual Information

作　　者：陈平华[1] 黄辉 麦淼[2] 周宏虹 Chen Ping-hua;Huang Hui;Mai Miao;Zhou Hong-hong(School of Computers,Guangdong University of Technology,Guangzhou 510006,China;Guangdong Nanfang Media Group,Guangzhou 510601,China;Guangdong Science and Technology Innovation Monitoring and Research Center,Guangzhou 510033,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006 [2]广东南方报业传媒集团有限公司,广东广州510601 [3]广东省科技创新监测研究中心,广东广州510033

出　　处：《广东工业大学学报》2018年第5期20-25,50,共7页Journal of Guangdong University of Technology

基　　金：国家自然科学基金资助项目(61572144);广东省科技计划项目(2013B091300009;2014B070706007;2017B030307002)

摘　　要：针对传统单标签特征选择算法不能直接应用于多标签数据的问题,提出一种多标签特征选择算法——MMLRF算法.在ReliefF的基础上,MML-RF算法提出新的类内最近邻样本查找方式,并结合多标签的贡献值改进特征权值的计算方法,能很好地适应多标签数据的特点;同时为了减少特征冗余,MML-RF算法以互信息作为特征冗余度量方式,提出一种去冗余方法,能够得到更小的特征子集.实验表明,MML-RF多标签特征选择算法得到的特征子集规模较小,且在多标签数据集上具有很好的分类效果,能够提升多标签学习和数据挖掘工作的效率.In view of the problem that the traditional feature selection algorithm can not be applied to the multilabel learning context,a MML-RF algorithm is presented.The MML-RF improves the way of defining and searching nearest neighbor on the basis of the ReliefF,and introduces a new parameter to consider the contribution values of different labels.The improved weighting formula enables MML-RF to be used to the multi-label dataset.MML-RF algorithm makes use of mutual information as the measure of feature redundancy,and puts forward a solution to redundancy,which can get smaller subset of features.Experiments show that the feature subset of MML-RF is smaller,and has good classification effect on multi-label dataset,which can further enhance the efficiency of subsequent multi-label learning and data mining.

关 键 词：特征选择 多标签学习 RELIEFF 互信息 特征冗余 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]