Multi-label Learning With Limited Or Missing Labels

Traditional single-label learning assumes that each point only belongs to one category,but with the development of annotation technology,each point may be assigned to multiple different categories simultaneously,i.e.,multi-label data.In real applications,it is difficult to acquire the label of multi-label data.Generally,the obtained labeled data is limited or incomplete.For example,in image annotation,human experts have to go through the entire list of all candidate words in order to decide the set of all possible labels for an image.It requires excessive resources to manually label each image with all its labels,and generally,we only obtain a limited data or partial labels for data points.In the case of limited labeled data or missing labels,we put forward related learning methods in two cases,respectively.With the limited of labeled data and high-dimensional data,the method Semi-Supervised Multi-Label Feature Learning via Label Enlarged Discriminant Analysis is proposed.With missing labels,this paper presents a method Low Rank Multi-Label Classification with Missing Labels.The three main contributions of this paper are as follows:(1)In this paper,we have analyzed and researched multi-label classification with insufficient labels.Combining with the characteristics of insufficient labels and demands of practical applications,on the base of the relevant literatures,respectively,multi-label dimensionality reduction method with limited labeled data and multi-label classification with missing labels are proposed.(2)In the case of limited and high-dimensional labeled data,the method LEDA(Semi-Supervised Multi-Label Feature Learning via Label Enlarged Discriminant Analysis)is proposed.LEDA algorithm propagates the label information from the labeled data to the unlabeled data through a designed multi-label label propagation method.It then incorporates the enlarged multi-label information to learn a transformation matrix for dimensionality reduction.In this way,the information of the labeled data and unlabeled data can be explored more effectively to learn a better subspace,which is beneficial to the classification.Theoretical analysis and experimental analysis verify the effectiveness of the proposed method.(3)In the case of missing labels,the method LRML(Low Rank multi-label Classification with Missing Labels)is proposed.LRML algorithm joints the underlying true label matrix recovery and multi-label classifier learning.With the assumption of label consistency and local invariance,the information of features is embedded in the label space.Then,the correlation among labels and robustness are characterized by low rank and regularization constraint.Theoretical analysis and extensive experiments show the effectiveness of the proposed algorithm.The experimental results demonstrate that our proposed algorithm can achieve superiority compared with other methods.