Correlation-Based Weighted K-Labelsets for Multi-label Classification

摘要：

　　RAkEL(RAndom k-labELsets) is an effective ensemble multi-label classification method where each sub classifier is trained on a small randomly-selected subset of k labels,called k-labelset.However,random combination of labels may lead to the poor performance of sub classifiers and the method can not make full use of the label correlations.In this paper,we propose a novel ensemble multi-label classification method named LCWkEL(Label Correlations-based Weighted klabELsets).Instead of randomly choosing subsets,we select a number of klabelsets based on a label correlation matrix.Furthermore,considering the label correlations in different k-labelsets may have different influence on an instance,we construct a weight coefficient vector for an instance.Each dimension of the vector represents the weight coefficient for each sub classifier.For themulti-label classification of an unlabeled instance,LCWkEL calculates the weighted sum of all sub classifierspredictions,which can improve the classification performance effectively.Experimental results on three areas of data sets show that the method proposed in this paper can obtain competitive performance compared with the RAkEL method and other high-performing multi-label classification methods.

关键词： Multi-label classification Label correlation Ensemble method k-labelsets

作者: Jingyang Xu Jun Ma

作者单位: School of Computer Science and Technology,Shandong University,Jinan 250101,China

会议类型: 国际会议

会议名称: International Asia-Pacific Web Conference(第18届国际亚太互联网大会)

会议地点: 苏州

会议语种:英文

页码: 408-419

在线出版日期: 2016-09-23（万方平台首次上网日期，不代表论文的发表时间）