Research On Multi-view Subspace Clustering Methods

With the rapid development of the Internet and other technologies,a large amount of data has been generated as a result.Clustering as a classical data mining method has been widely used,and its different branches have been developed rapidly.Among them,multi-view subspace clustering method is proposed for the high dimensionality and multiple features of existing data,which has become a research hotspot in recent years.In this paper,we optimize the existing multi-view subspace clustering methods from the perspectives of specific subspace learning,multi-view consistency,and difference mining,and extend the proposed methods to a practical application-lung nodule segmentation.The specific work is as follows.1.Research on Multi-View Subspace Clustering Method with Joint Tensor Representation and Instruction Matrix Learning.To address the problem that most current multi-view subspace clustering methods focus on local inter-view connections and ignore the higher-order associations between views,we propose a Multi-View Subspace Clustering with joint Tensor representation and Indicator matrix learning(MVSCTI).In the framework based on sparse representation,this method uses Tensor Singular Value Decomposition(T-SVD)to impose low-rank constraints on a tensor stacked by multiple view similarity matrices to mine the correlation between views efficiently,as well as to incorporate the subsequent spectral clustering into the MVSCTI model,so that the similarity matrix and indication matrices can be jointly optimized to facilitate the acquisition of the global optimal solution.The experimental results demonstrate the effectiveness of joint tensor representation and indication matrix learning.2.Research on Multi-view Subspace Clustering with Consistency and Complementarity Fully Mined.To tackle the problem that multi-view information has not been fully mined,a Multi-view Subspace Clustering with Consistency and Complementarity Fully Mined(MSC~3FM)method is proposed.In order to fully mine the multi-view information,the method uses T-SVD to impose low-rank constraints on the tensor composed of multiple view similarity matrices which ensures the global low-rank structure;by increasing the difference of similarity matrices between views through orthogonality constraints to ensure the mining of local complementary information;by introducing the consistency constraint term to retain the consistency information between views.In addition,considering the variability of the contribution of different views,a weighted Multi-view Subspace Clustering with Consistency and Complementarity Fully Mined(WMSC~3FM)method is proposed on the basis of MSC3FM.Fully Mined(WMSC~3FM),which designs an adaptive weighting strategy to assign different weights to different views.The experimental results demonstrate that the full consideration of discrepancy and consistency helps to mine the multi-view information.3.Research on lung nodule segmentation method based on multi-view subspace clustering.In order to explore the applicability of the multi-view subspace clustering method and extend its application,and considering the current demand for intelligent assisted medical diagnosis,we applied it to lung nodule segmentation and proposed a multi-view subspace clustering-based lung nodule segmentation method.In order to ensure the segmentation effect,the lung CT images are pre-processed before input to the multi-view subspace clustering method,including two parts of noise removal and region of interest extraction;after that,the region of interest is input to the multi-view subspace clustering method,and considering the low dimensionality of the pixel point-based features,the method connects the feature matrices extracted from different features in series,constructs the similarity matrix using Euclidean distance,and finally uses graph cut to obtain the final segmentation results.The feasibility and effectiveness of the multi-view subspace clustering method for lung nodule segmentation are confirmed by the segmentation experiments on the actual lung nodule dataset.