| With the development of information acquisition technology,the complexity and universal dimension of data are increasing day by day.Multi-view data with different expression forms appear in many practical applications of big data mining and analysis.These multi-view data come from different information sources and have different statistical characteristics,which can better capture the heterogeneous complementary information from different data perspectives.The traditional clustering research for single-view data is gradually approaching the bottleneck because it cannot fully reflect the hidden structural information in multi-view data.Therefore,it is particularly important to explore the hidden value of information by integrating heterogeneous and complementary information to develop unsupervised multi-view clustering methods.This paper studies multi-view projection clustering from many aspects and is committed to designing novel multi-view clustering algorithms.The main research contents are as follows:(1)Aiming at the problems that most existing multi-view clustering algorithms do not fully consider the impact of noise and redundant information in the original high-dimensional data on the clustering effect,and usually rely on the fixed graph similarity matrix based on all views to optimize the two-step clustering strategy of the target,a dynamic fusion multi-view projection clustering algorithm is proposed.The algorithm integrates adaptive reduceddimension graph learning,parameter-free self-weight graph fusion and spectral clustering in the same framework,and the three processes promote each other.The rank constraint is applied to the Laplace matrix of the consensus matrix obtained in the dynamic fusion process to directly obtain the clustering results.In addition,the algorithm realizes parameter adaptation and designs an effective alternating iteration method to solve the resulting joint optimization problem.The algorithm has achieved good results on different artificial data sets and real data sets.(2)Aiming at the problems that some traditional multi-view clustering algorithms do not make full use of the node relationship of the original high-dimensional data,and ignore the diversity of the graph,which leads to the inability to separate into subspaces,a multi-view projection clustering algorithm based on graph filtering is proposed.The algorithm integrates dimensionality reduction subspace learning,global similarity learning and spectral clustering in the same framework,and jointly optimizes the projection matrix,potential space matrix,global similarity matrix and clustering label.Based on the assumption that different views have the same potential representation,we project the high-dimensional data into the potential low-dimensional subspace and establish a global similarity matrix on the potential space.In addition,the algorithm applies the graph filtering technology to inject the graph similarity into the data features,so as to learn the latent space matrix in the smooth representation,and designs an effective alternating iteration method to solve the joint optimization problem.Experimental results on several widely used real data sets show the superiority of the algorithm.(3)The two multi-view clustering algorithms proposed in this paper are applied to the field of face clustering in video.First,the face detector is used to obtain the initial face data set.After the faces are detected,they are aligned and the image features of each face are extracted to obtain a multi-view dataset suitable for the multi-view clustering algorithm.Finally,the algorithms proposed in this paper are applied to cluster video face images to obtain the clustering results,and four clustering evaluation indices are used to evaluate the clustering performance.The results show that the two algorithms proposed have certain practicability. |
