Missing Multi-label Learning For Label Semantic Space Mining

In multi-label learning,each instance corresponds to multiple labels and has rich semantic information.It is undeniable that the default phenomenon of labels often occurs in multi-label data sets.The default label will cause the multi-label learning algorithm to obtain wrong semantic information or lose important semantic information.This makes the work of label recovery particularly important.The mining of label space is a research hotspot in multi-label learning,and the obtained semantic information can improve the performance of the algorithm.Therefore,the method of mining label space and the method of measuring information are the research focus.Based on this,the research work of this article on the above issues is as follows:1)In multi-label algorithms,feature and label embedding are wildly used to mine the semantic information of the label space.However,these methods do not take advantage of the possible correlation information between features and labels.The proposed label-specific features better interprets the relationship between the feature and the label,the label may correspond to a set of its own features.However,this type of method fails to give a logical relationship between the feature and the label,and whether the label and the instance may have the same logical relationship.Therefore,this thesis proposes a Multi-label classification algorithm based on PLSA(Probabilistic Latent Semantic Analysis)learning probability distribution semantic information.2)At present,most of the algorithms combine label correlations and label-specific features to improve the multi-label learning effect,but do not consider the impact of label marking errors or defaults in data sets.In fact,the label completion method can further enrich the information of label matrix,and then the joint learning framework of joint label-specific features can effectively improve the robustness of the multi-label learning algorithm.Based on this,this thesis proposes a multi-label learning algorithm for joint label completion and label-specific features for multi-Label learning algorithm.3)The missing label data will lead to the label embedding model capturing incomplete inherent information.However,the label correlations recovery mechanism only considers the label correlations but ignores the objective existence of instancecorrelation information.Therefore,we propose a two-level label recovery-based label embedding for multi-label classification with missing labels.Based on PLSA,this thesis makes a reasonable explanation of the relationship between label-specific features and labels and instances.In this thesis,a joint learning framework of label completion and label-specific features is constructed to solve the influence of default label on the extraction of generic attributes label-specific features.For the problem of obtaining the internal information of the default label space,this thesis uses the two-level label recovery mechanism and the method of label embedding.The algorithm proposed in this thesis has better performance with other comparison algorithms.Visual analysis and statistical hypothesis testing are used to further illustrate the rationality of the proposed algorithm.