Research On The Multi-view Clustering Algorithm

More and more samples can be represented by different features of multiple views.With the increasing diversity and dimension of data,the research on single view clustering algorithm is close to the bottleneck.Based on the assumption that features complement each other,multi-view clustering can effectively combine heterogeneous features and divide data points into different categories,so as to obtain better clustering.This paper studies multi-view clustering from multiple perspectives and improves the multi-view clustering algorithm.The specific research is as follows:(1)A multi-view clustering algorithm based on Grassmann manifold fusion subspaces was proposed.When learning subspace,the existing multi-view clustering algorithms assume that there is a linear relationship between multi-view data points,or the locality of the original feature space cannot be preserved in the process of learning.In addition,the direct fusion of subspace representations of different views in Euclidean space is too rigid to align the learned subspace representations.The multi-view clustering algorithm based on the fusion subspace of Grassmann manifold can effectively solve the above problems.The algorithm consists of the following three parts: one is to combine kernel learning and local manifold structure learning to obtain the subspace representation of different views,kernel learning can map the nonlinear data in the original feature space to the high-dimensional kernel space and the local manifold structure makes the data similar to the original feature space similar in the representation space;the second is to fuse these subspace representations on the Grassmann manifold to obtain the consistent affinity matrix,so as to promote the alignment of the subspace representation matrix corresponding to each view;the third is to impose rank constraints on the consistent affinity matrix,so that the number of connected components of the consistent affinity matrix is equal to the number of clusters,so as to promote the direct division of samples into clusters.These three parts are integrated into a framework,and the alternating direction multiplier method is used to solve the constrained optimization problem.The algorithm has achieved good results in 8 benchmark data sets.(2)A multi-view clustering algorithm based on partition fusion and tensor representation is proposed.The existing multi-view clustering algorithms ignore the noise of multi-view data and the inconsistency between the heterogeneous features of multi-view when integrating multi-view information.When fusing multi-view information,they treat each view equally,can’t fully consider the weight difference of different views,or ignore the high-order related information of multi-view in the process of learning.The multi-view clustering algorithm based on partition fusion and tensor representation can effectively solve the above problems.The algorithm is divided into two steps: First,the subspace representation of each view is obtained by combining kernel learning and local manifold structure,and the importance of different views is measured by adaptive weight learning;Secondly,the corresponding partition space is generated according to the subspace representation;Thirdly,the partition matrix is reorganized into a third-order tensor,and the tensor nuclear norm is applied to capture the high-order correlation of multi-view clustering.The partition space learning and tensor low-rank representation are unified into one framework,and an effective optimization algorithm is designed to solve the optimization problem.The algorithm has achieved better clustering results on 12 benchmark datasets.(3)Air quality is an important basis for determining a city’s development.Multi-view clustering algorithm can be applied to air quality classification.Firstly,analyze the factors affecting air quality,collect relevant data,and preprocess the data to obtain multi-view dataset.Then,the two multi-view algorithms proposed in this paper are applied to the dataset for air quality analysis.Compared with 11 comparison algorithms,the experimental results show that the algorithm has good performance and certain practical value.