Multiview Subspace Clustering Based On Linear Representation

In the real world,the data that people observe from different viewing angles,generate from different sources and combine various features is called multi-view data.Multi-view learning uses the complementarity and consistency of information from multiple perspectives of data view to realize information fusion and provide support for rational decision making.Recent years,multi-view learning has become a research hotspot in pattern recognition,image processing,natural language processing and other fields.Multi-view clustering based on linear representation has been widely concerned by researchers because of its characteristics such as easy implementation,small computation and good clustering performance.However,due to the heterogeneity of inter-view information,feature redundancy and noise,it is difficult for the existing multi-view clustering methods based on linear representation to learn a consistent representation matrix from multi-view data that can effectively reveal the underlying structure of multi-view data.Therefore,this thesis,starting from the three aspects of affinity matrix noise,feature learning pattern and view fusion method,deeply studies the more robust consistent representation multiview clustering algorithm,and designs three multi-view clustering methods,the specific contents of which are summarized as follows:(1)Nested Low-Rank Representation Structure for Multi-view Subspace ClusteringThe affinity matrix learned by traditional single-structure methods is often accompanied by a large amount of noise,which will result in suboptimal clustering performance.In this thesis,the Nested Structure Low-Rank Representation multi-view subspace clustering method(NESLRR)is proposed.The method incorporates Least Squard Regression(LSR)and Low-Rank Representation(LRR)into a unified algorithm framework.The uniform representation matrix is learned from the joint multi-view data after Feature Concatenation(FC).This not only takes full use of the diversity of multi-view information,but also learns a more reasonable and clean affinity matrix.In addition,we design an Augmented Lagrangian Multiplier(ALM)algorithm to solve the objective function.The simulation experiments are carried out on five real world data-sets,such as face,text and scene,and the experimental results show that the clustering performance of the proposed method is improved.(2)Adaptive Weight Structure for Multi-view Subspace ClusteringIn the process of multi-view learning,the features are often treated equally,and the redundant features of the views in the learning process will have a negative impact on the consistent representation,so that the best clustering effect cannot be achieved.Therefore,this paper proposes adaptive weight structure for multi-view subspace clustering.The adaptive weight matrix is designed to dynamically adjust the error matrix in the learning process of view representation,and small weights are applied to large corruption values,and vice versa,so that the learned affinity matrix is cleaner and more reasonable.In addition,we design an Augmented Lagrangian Multiplier(ALM)algorithm to solve the objective function.Simulation experiments are carried out on five real world data sets,such as face,text and scene,and the experimental results show that the proposed method improves the clustering performance.(3)Joint Similarity Representation for Multi-view subspace clusteringMost of the traditional multi-view clustering algorithms adopt the consistency method with similarity strategy to obtain the consistent representation of views.Although the consistency method is helpful to explore the underlying structure of multiple views,due to the different statistical characteristics of different views,the corresponding view representation structure presents its own unique structural characteristics.The redundancy and noise in the individual view representation structure can irreversibly negatively affect the final consistent representation and result in suboptimal clustering performance.To this end,this paper proposes Joint Similarity Representation for Muti-view Subspace Clustering(MSC-JSR).It not only preserves the unique structural characteristics of each view in multiple views,but also avoids the negative impact of inherent structural differences between views on the final consistent representation.The simulation experiments are carried out on eight data-sets,such as face,text,number,object and scene,and the experimental results show that the clustering performance of the proposed method is improved.