Diversity Representation For Multi-view Multi-label Learning

With the rapid development of information technology,big data has penetrated into every aspect of life,and the manifestations of data have gradually shown richness and diversity.Each object has rich semantic information and diverse representations.Multi-view multi-label learning is a basic framework for studying this type of problem,and it has become one of the research hotspots in the field of machine learning due to the widespread existence of multi-view multi-label data.In recent years,although some progress has been made in multi-view multi-label learning tasks,there are still some problems.On the one hand,as dimension differences between different views may be large,some existing methods extract representations into the same dimension,which may affect the model performance.And how to better capture label correlation is also a difficult point.On the other hand,the previous algorithms have not considered the sample correlation,and the process of exploring the consistency and complementarity between multiple views is through linear mapping and adding some constrains,without explicit mining.Therefore,based on the above two problems,this paper starts research and proposes two methods,which are specifically introduced as follows:1.Label correlation induced multi-view multi-label learning.This method can easily set the dimensions of different representations and induce the extraction of representations through label correlation.Firstly,the independence of the private representations is constrained by adding mask value to the weight value of the network,and the independence of the shared representation and each private representation is constrained.Then,the label correlation is explored by the the word embedding of labels and adaptive graph convolution neural network,and the correlation is integrated into the process of representations extraction to make the representation more discriminant.2.Graph based commonality and individuality representations for multi-view multi-label learning.This method incorporates sample correlation through graph structure,and uses graph convolutional neural network to explicitly explore the consistency and complementarity in multiple views.Firstly,the intermediate representations are extracted from different views and the graph structures are constructed.The first K large value is selected by kernel function as the adjacency matrix.Through different graph convolutional neural networks and the intersection of different adjacency matrices,the commonality of all views is explored,and the individuality of different view is explored based on the same network parameters and their respective adjacency matrices.