Research On Weighted Multi-view Subspace Clustering Algorithm

With the emergence and development of big data,data acquisition and storage technologies have been increasingly enriched.There are lots of collected data that are formed with multiple sets of features,i.e.multi-view data.For example,images of the same object obtained from different angles are multi-view data.This kind of data contains features that describe a specific object from different directions,thus including a great amount of complementary information as well as redundant information.These features usually are described with high dimensions.A significant challenge has been posed that how can we analyze these multi-view data efficiently and accurately.Clustering is one of the basic solutions,which has been highly concerned and developed rapidly.Among those clustering methods,one major direction of solving multi-view data clustering is multi-view clustering on subspace.This paper focuses on how to effectively use the correlation and diversity among views for multi-view subspace clustering.The optimization methods for learning the weights of views in multi-view subspace clustering by exploring the allocation of weights among multiple views and utilizing relationships between views,which finally proves to enhance the overall performance of clustering.The main research contents of this paper include:(1)Studying on spectral clustering algorithm.Spectral clustering is the basis of sparse subspace clustering and also the foundation of this paper.This paper first introduces the theoretical background of spectral clustering algorithm,classical spectral partition method,graph theory,and Laplacian matrix.However,spectral clustering requires k-means as a post-processing method for clustering,which is sensitive to initialization.So,the graph clustering algorithm with Laplace rank constraint is introduced to solve it.In this paper,graph clustering with Laplace rank constraint is applied to the multi-view subspace clustering instead of other traditional spectral clustering algorithms,aiming at solving the problem of instability caused by the sensitivity of k-means to initialization.This algorithm can directly obtain the clustering index through similar matrix,without k-means or other post-processing methods.(2)Designing of self-weighted multi-view subspace clustering algorithm.In order to reasonably make use of the correlation and diversity among multiple views,a weighting method for multi-view clustering is studied,based on the weights of parameters and low-rank tensor constraint.And a new multi-view subspace clustering algorithm based on self-weighting is then proposed by further combining the self-weighting with low-rank tensor constrained multi-view clustering algorithm.The method proposed not only achieves the goal of capturing high-dimensional correlation of multi-view data,but also allocates weights appropriately according to their different contribution to clustering.Experiments have proved that the proposed algorithm effectively can improve the performance of multi-view subspace clustering.(3)Studying on dynamic allocation of reasonable weights in multi-view subspace clustering.Based on different contribution of different views to clustering performance,an entropy-based multi-view weighting method is proposed by studying more appropriate weighting methods for allocating weights of different views.In order to capture high-dimensional correlation of multi-view data and simultaneously avoid the influence of k-means,an extended multi-view subspace clustering method based on Laplace rank constraint of L1 norm and L2 norm distance and low-rank tensor constraint is introduced in this paper.The entropy-based weighted multi-view subspace clustering algorithm based on entropy can effectively achieve dynamic allocation of reasonable weights by utilizing the high-dimensional correlation and diversity among different views.