Multi-view Clustering Based On Matrix Decomposition

With the development of the Internet,data information has become an indispensable part of people’s production and life.The massive amount of data generated in the information age will show the phenomenon of high dimensions and multiple views,how to efficiently obtain usable information in complex massive data has become one of the main research directions.As an effective high-latitude and multi-view data analysis method,multi-view clustering uses the characteristics of the data itself to train reasonable data processing models without any label information,and has been widely used in data mining,data analysis and other fields.Among the existing multi-view clustering methods based on matrix factorization has achieved good performance,but it still exists but there are still some problems that need to be solved urgently:(1)Only considering the first-order neighborhood relationship of the data,the high-order relationship of the data is not fully utilized,and the learned data neighborhood structure is not optimal;(2)The diversity information is not mined,only the consistent information between the data is considered,and the complementary information of the data is not fully utilized;(3)The low-dimensional representation and clustering of data are two independent processes,and it is difficult to achieve the optimal of both at the same time without information interactive learning;(4)Considering only the ideal situation of complete views,it is impossible to cluster the missing view data that is more common in daily life,that leads to the lacking of the practical application of clustering technology.To solve the above four problems,this paper put forward to three multi-view learning methods based on matrix factorization:(1)This paper put forward a Robust Multi-view Non-negative Matrix Factorization for Clustering(RMNMF)method,which can learn the second-order neighborhood information of the data on the basis of the first-order neighborhood of the data,and obtain the optimal neighborhood information of the data through the optimal combination of first-order information and second-order information.In addition,by introducing the Hilbert-Schmidt independence criterion(HSIC),the diversity of information between different views is mined to make full use of complementary information.Experiments on commonly used multi-view datasets show that the new method proposed in this paper has superior performance compared with the clustering methods.(2)This paper proposes a Virtual Label Guided Multi-view Non-negative Matrix Factorization for Data Clustering(VLMNMF)model,which integrates the latent representation learning and cluster learning of data into an overall learning framework.Through interactive learning,clustering guides the learning of the latent representation,while the low-dimensional representation learning of the data also improves the accuracy of clustering.In addition,by learning virtual labels,the latent representation matrix of data can be more discriminative.Experiments on commonly used multi-view datasets demonstrate that the new model proposed in this paper has superior performance compared with the existing multi-view clustering methods.(3)This paper proposes an Incomplete Multi-view Clustering via Virtual-label Guided Matrix Factorization(VLMF)model,which broadens the virtual label information and one-step clustering into incomplete view data,which enhances the application of clustering and ensures the clustering accuracy.In addition,the model improves the local geometry mining of the data,and proposes a parameter-free model which reduces the pressure of parameter selection while achieving full mining of data information.Experiments on commonly used multi-view datasets show that the new model presented in this paper has superior performance compared with the existing incomplete multi-view clustering algorithms.