A Study Of Multi-view Learning

With the continuous development of the Internet,all kinds of behavior in life produce a large amount of data.It is often necessary to deal with multi-view data from various sources of information or to describe things from many aspects,and the study of multi-view learning methods for such special data came into being.The existing multi view learning method cannot make full use of the single view and the relationship between views.In this paper,we study three multi-view learning problems of multi-view reduction,multi-view clustering and multi-task multi-view clustering.First of all,we study the semi-supervised dimensionality reduction,hybrid structure dimension reduction and multi-source multi-view learning algorithm.By adding the hybrid structure of preserving the structure of the sample structure and the discrepancy penalty,multi-view semi-supervised dimensionality reduction algorithm is improved.The low dimensional representation for each view which obtained by semi-supervised dimensionality reduction algorithm does not make use of single view information,and ignore the potential structure and local structure of the samples.Therefore,the hybrid structure of preserving the structure of the sample structure is introduced.It can reduce the loss of information because of dimensionality reduction.Then the discrepancy penalty in multi-view learning is introduced to dig the consistency of the multi-view,so as to combine results of multiple views.After the new algorithm is executed,the clustering performance of the reduced dimension samples is improved.Secondly,we study the sparse subspace clustering based on spectral clustering and the related multi-view clustering algorithms.By modifying the multi-view data into the same low-dimensional space and adding the manifold regular term to preserve the local structure,Multi-view sparse subspace clustering algorithm is improved.On the basis of the original algorithm,by analyzing the characteristics of multi-view sharing the same sparse representation matrix,we consider that multi-view data is mapped from the original high-dimensional space to the same low-dimensional space,mining the views information.At the same time,in order to keep local structure of the original feature space,manifold regularization is introduced.Sparse representation of the samples is obtained by the low dimensional space,which can better reflect the information of the samples.This improvement makes the clustering performance of the algorithm improve.Finally,we study the multi-task multi-view clustering problem and its related multi-view learning algorithm.We improve the bipartite graph based multi-task multi-view clustering by introducing the augmented view and eliminating the difference minimum constraint.Because multi-view data share the potential clustering structure,the potential multi-view clustering structure is used to directly eliminate the difference of multi-view clustering structure of single task.In each task,clustering indicator matrix of a single view can be replaced with the same matrix,which emphasizes consistency between views.Furthermore,the augmented view generated by multiple views is introduced.We use the rich structure information of the augmented view to modify clustering structure of single task to compensate for the loss of information,further to mine view information and improve performance.In the improved algorithm,the average clustering performance of multiple tasks is improved.The above improved algorithms have been carried out on a number of public data sets to prove the effectiveness of the algorithms.