Adaptive Label Correlation And Instance Correlation Guided Incomplete Multi-View Weak Label Learning

In real life,objects usually have different view representations and rich semantics.As a paradigm for analyzing such data,multi-view multi-label learning has attracted much attention in recent years and has been widely used in many fields,such as heterogeneous multimedia data analysis and bioinformatics.For the problem of multi-view multi-label learning,there are many mature methods.However,most of these methods are based on the assumption of complete data.In practical applications,due to incomplete view data collection and the limitations of manual annotation,data is inevitably incomplete,which will greatly increase the difficulty of multi-view multi-label learning.Without loss of generality,in this study,for a certain sample,we define incomplete views as the complete absence of partial views that characterize the instance,incomplete labels as uncertainty about whether some labels are associated with the instance,and the method to solve the problem of incomplete views and incomplete labels in multi-view multi-label learning is defined as incomplete multi-view weak label learning.In response to the above challenges,this paper proposes two models from different perspectives.The main research contents are as follows:(1)Focus on the problem that both incomplete view and incomplete label in multi-view multi-label learning,the model called shallow incomplete multi-view weak label learning is proposed.Firstly,the model assumes each view of the instance is obtained from a latent shared representation through different mappings,and matrix factorization is used jointly with the indicator matrix to make full use of the existing incomplete multi-view weak label data.Then,the technology of learning the standard Laplacian matrix in graph theory is introduced to describe the label correlation and instance correlation.Thereby embedding manifold regularization in the model to make the learned latent shared representation and classifier more reasonable.Finally,the experimental results on multi-view multi-label datasets show that the proposed model can well solve the problem of incomplete views and incomplete labels in multi-view multi-label data.(2)Focus on the limitation of shallow linear model in mining complex nonlinear relationships,the model called deep incomplete multi-view weak label learning is proposed.Firstly,we use degenerate networks to reconstruct each view from the latent shared representation,so the latent shared representation can encode complete information from multiple views,and indicator matrix is used to handle the problem of incomplete view and incomplete label.In addition,we embed the label correlation and instance correlation in the model through the technique of learning standard Laplacian matrices to make the learned latent shared representation and prediction network are more reasonable.Finally,the experimental results on datasets show that the deep incomplete multi-view weak label learning model can further improve the performance of the algorithm on the basis of the shallow incomplete multi-view weak label learning model.