Research On Robust Multi-view Subspace Clustering Algorithm

In the information age,data information has shown an explosive growth trend,which in turn has produced multi-view data with multiple feature descriptions.In the face of massive data,it is important to classify the data effectively according to similarity to reduce the clutter of data,so as to better help scholars analyze and utilize the potential value of data.Multi-view clustering methods can characterize the data from different perspectives,effectively reveal the internal structure of the data,and usually have better clustering performance than single-view clustering methods,so their application in multi-view data analysis is becoming more and more widespread.Currently,subspace clustering shows good performance in handling multi-view data clustering tasks.There are two key problems in processing multi-view data using subspace techniques:(1)many existing algorithms assume that all views have the same representation matrix and do not fully utilize the underlying information of multi-view data,since the representation matrix of different views should have the same clustering properties,rather than being consistent across multiple views;(2)Many existing algorithms make use of the fixed similarity matrix of all views for clustering,and do not fully consider the differences among different views.In view of the above problems,this thesis conducts research from the following two aspects:(1)Low-rank multi-view subspace clustering algorithm based on orthogonal constraints(LRMVOC).LRMVOC first learns a low-dimensional feature subspace using the projection matrix to reduce the impact of noise and redundant information on the subspace representation learning.The algorithm then decomposes the representation matrix obtained in the lowdimensional feature subspace into view-specific basis matrices and shared encoding matrices,while imposing orthogonality constraints on the basis matrix of each view,and using kernel parametrization to ensure the low-rankness of the shared encoding matrix.The algorithm uses an improved matrix decomposition method for learning multi-view complementary information and consistency information,which is beneficial to improve the clustering effect of the algorithm.Experiments on five image datasets validate the rationality of the proposed LRMVOC algorithm.(2)Robust multi-view subspace clustering with consistency graph learning(RMCGL).RMCGL first learns a potential representation matrix for each view,and subsequently learns a similarity matrix for the corresponding view on top of it.Finally,a unified consistent graph matrix is learned.RMCGL uses a self-weighting approach to assign reasonable weights to each view.The advantage of this algorithm is that it fully considers the complementary information between different views and utilizes a unified optimization framework for consistent graph learning and algorithm optimization in the latent subspace.By imposing rank constraints on the consistent graph,the consistent graph has an optimal clustering structure,and the clustering results are obtained directly on the consistent graph.In this thesis,we design several experiments to effectively verify the clustering effect as well as the generalization ability of the RMCGL algorithm.