Research On Multi-view Subspace Clustering And Community Detection Based On Robust Subspace Representation

With the rapid development of science and technology,people's ability to acquire data is becoming stronger and stronger.The acquired data is often characterized by large scale,high dimension,multi-view and diversity.Therefore,it is particularly important to analyze meaningful and valuable information from the data.The analysis of the data needs to reveal the inherent complexity and real dimensions of the data and cover the global and local structural features of the data.The subspace representation algorithm is an effective method to deal with high-dimensional data because it is able to mine potential low-dimensional data structure features from high-dimensional data.In recent years,the problems related to subspace representation have become popular among scholars.For example,the subspace representation algorithm based on sparse representation,low rank representation and Frobenius norm constraint has achieved certain efficient performance in the analysis of high-dimensional and multi-view data.However,there are still some problems in this algorithm,such as insufficient data mining and application.For multi-view data,the existing multi-view subspace clustering algorithm still has shortcomings when coordinating and utilizing the data information from all views,resulting in the poor performance of the algorithm.In addition,the community network data is also a kind of complex high-dimensional data.In this paper,the subspace representation is further applied to the analysis of social network data.The nodes in each community can be regarded as a subspace in the geodesic space.At present,the research on community detection algorithm based on subspace representation is still insufficient,while the existing related algorithms are very sensitive to the noise in the network and lack the ability to accurately learn the community structure.To solve the above problems,the work of this paper is mainly from the following aspects:(1)A robust subspace clustering algorithm based on adaptive parameter learning is proposed.Based on the robust subspace representation of Frobenius norm constraint,we studied the sparse representation and the affinity matrix at the same time.Importantly,we set one of the parameters to be acquired adaptively without manual setting.For unsupervised clustering algorithm,the less parameters are,the more robust the algorithm is.Although the adaptive learning of a parameter may achieve the same effect by exhausting more parameter values,when experiments are conducted on multiple data sets,the efficiency will be greatly reduced if the iterative exhausting is done every time.(2)The learning of robust subspace with adaptive parameters is extended to a multi-view algorithm model.The adaptive parameter setting is of great significance to unsupervised clustering algorithm.In many traditional multi-view clustering algorithms,the parameters are often set to a unified value when learning data from each view,which may make the algorithm miss the best parameters.If the parameters are set by the exhaustive method,the number of experimental parameters will increase with the increase of the number of data view.Therefore,for the multi-view subspace clustering algorithm proposed in this paper,one of the parameters will be self-adaptive learned from the all data views,which not only greatly reduces the number of parameter adjustment experiments,but also ensures that the optimal parameters are learned.(3)The robust subspace representation of adaptive parameter learning is extended to community detection.The community in the network can be represented by nodes in the same community,in order to find the nodes contained in each community,we fuse community detection and robust subspace representation algorithm to learn subspace,and learned subspace is equivalent to distinguishing the structure of network community.Compared with the existing community detection algorithms based on subspace representation,the algorithm of parameter is set to adaptive learning,moreover,the algorithm to learn the global and local manifold structure of network data,so it is more robust to the noise in the data,more depth analysis of the data,and better learning of the data.In addition,on the synthetic data sets and the real-world data sets,the experiments have proved its excellent performance.