摘要
现有的多标记降维算法常通过学习标记相关性构建样本间的相似关系,进而提高学习系统的性能.然而,在实际应用中,样本的标记信息可能存在噪声,且部分标记信息可能缺失,因此由样本的标记信息学得的标记相关性可能不准确,无法有效挖掘样本间的相似关系.为了解决该问题,从样本的特征空间与标记空间两个方面构建样本间的相似关系.在利用标记空间学习标记相关性的同时,通过引入特征空间中的概率超图模型,提出一种嵌入样本流形结构与标记相关性的多标记降维算法.在十个多标记数据集和六种评价准则上的实验结果证明了所提算法的有效性.
The existing multi-label dimensionality reduction algorithm often constructs the similarity relationship among samples by learning the correlation between labels to improve the performance of the learning system.However,in practical applications,the label information of samples may be noisy,and some of the label information may be missing.Therefore,label correlation learned from label information of samples may be inaccurate,and the similarity relationship among samples cannot be excavated effectively.To solve this problem,this paper tries to construct similarity between samples from two aspects:feature space and label space.In particular,learning label correlation from label space,meanwhile,by introducing the probabilistic hypergraph model into the feature space,a multi-label dimensionality reduction algorithm embedded sample manifold structure and label correlation is proposed.Experimental results on ten multi-label datasets and six evaluation criterias demonstrate the effectiveness of the proposed algorithm.
作者
马宏亮
万建武
王洪元
Ma Hongliang;Wan Jianwu;Wang Hongyuan(School of Information Science and Engineering,Changzhou University,Changzhou,213164,China)
出处
《南京大学学报（自然科学版）》
CAS
CSCD
北大核心
2019年第1期92-101,共10页
Journal of Nanjing University（Natural Science）
基金
国家自然科学基金(61502058
61572085)
江苏省高校自然科学基金(15KJB520002)
关键词
多标记降维
标记相关性
流形结构
超图
multi-label dimensionality reduction
label correlation
manifold structure
hypergraph
