Consistent Auto-weighted Multi-view Subspace Clustering

With the development of the applications in computer vision,cluster analysis is always an active field in data mining.In the field of clustering,subspace clustering is always a hot topic in computer vision.But in the era of diverse information,only using the information form one single view cannot meet the practical demand.Hence,to overcome the limitation of the traditional methods that only use one single view,multi-view subspace clustering attracts much attention.Because many multi-view subspace clustering methods focus on the establishment of affinity matrix by self-representation,this thesis will address the problems about this kind of method.Nowadays,most multi-view subspace clustering methods collect the information from multiple views by considering the properties such as consistency.Although they effectively overcome the limitation of the single view methods,they still suffer from some problems.For example,when addressing consistency,previous work often treats the information from different views equally.However,the learning effect of different views is usually different,and treating each view equally may lead to the view which is difficult in feature extraction affecting the overall clustering effect.To address this problem,in this thesis,we propose a novel multi-view subspace clustering method,in which each view can automatically assign reasonable weight to obtain more appropriate consistency in the clustering process.Although there are difficulties in solving the model by simultaneously using the self-representation and the auto-weighting strategy,we successfully design a special updating scheme to overcome this difficulty and prove its convergence theoretically.In addition,we consider the sparsity and density of the representation matrix by using₁l-norm and F-norm to ensure the learning effect of each view.Finally,the effectiveness of the proposed algorithm is verified on three data sets.The experimental evaluation results show that the proposed consistent auto-weighted multi-view subspace clustering outperforms the state-of-the-art.