Research On Multi-Label Classification Algorithm Based On Multi-View

Multi-view multi-label learning can effectively deal with high-dimensional heterogeneous multi-semantic data classification problems and is widely used in various fields in real life,such as medical diagnosis,image recognition,and recommendation systems.In multi-view multilabel learning,there are inevitably problems of consistency among views,diversity,non-aligned views,and correlation among labels.However,the existing multi-view and multi-label learning methods do not fully consider these issues,which severely restricts the performance of the classifier and the generalization ability of the model.Therefore,the information fusion of view consistency and difference between views in multi-view and multi-label learning is one of the critical problems that need to be solved urgently.In addition,the inconsistency of viewcorresponding labels is often ignored in non-aligned view learning,so how to deal with this inconsistency is another critical issue that needs to be solved.This dissertation mainly focuses on the above problems to build corresponding multi-view multi-label learning models based on different classification problem scenarios.The main contributions are summarized as follows:1)Aiming at the problem that the existing multi-view multi-label classification methods are insufficient in mining shared subspace information and missing labels limit the model,this dissertation proposes a two-stage multi-view multi-label missing label classification method based on subspace learning.The two-stage step-by-step learning strategy of this method can fully combine the respective advantages of the two task models,realizing the mining of shared subspaces and effectively dealing with the multi-label classification problem of missing labels.Statistical hypothesis testing and performance analysis experiments verify that the method has better performance.2)Aiming at the problem of non-aligned view learning,this dissertation proposes a multiview multi-label classification method based on view consistency and diversity of neural networks.This method solves the problem of information consistency and diversity among views under a unified neural network framework.This method can effectively deal with the problem of non-aligned view learning and enhance the completion of the original label matrix through label correlation.This method has better generalization ability through statistical hypothesis testing analysis and performance evaluation.3)For the case of non-aligned views,considering that the observed information of each view is different,resulting in the inconsistency among the corresponding label and the view,a two-stage method based on view-specific labels and feature-label dependency maximization is proposed—non-aligned multi-view multi-label classification methods.The method adopts a two-stage approach to deal with the problem of view-specific labeling and classification learning,respectively.In the first stage,the sample smoothness and consistency assumptions are fully utilized to learn view-specific labels based on the topology information of each view.The second stage maximizes feature-label dependencies to mine the correlation information between view features and labels for classification.After sufficient experimental analysis,it is verified that the method can learn the consistency and diversity of information among views while effectively acquiring the specific labels of each view.4)Aiming at the view-specific label problem and label correlation problem in unaligned view learning,this dissertation proposes a multi-view multi-label classification method that combines view-specific label and low-rank label correlation learning.This method jointly learns label correlation learning and view-specific labels.It uses complex label-related information to enhance and complete the view-specific label matrix so that the analyzed viewspecific label matrix information is more accurate and effective,which improves the classifier’s performance.The performance analysis shows that the scheme has good performance in generalization and effectiveness.