Research On Incomplete Multi-view Clustering Methods

Incomplete Multi-view Clustering(IMC)has attracted increasing research attentions due to its superiority in partitioning unlabeled multi-view data with missing instances.Although lots of progresses have been made,most existing IMC methods still have at least one of the following limitations:(1)The common relations among data points across all views and the complementary multi-view information of original data feature representation cannot be exploit simultaneously.(2)The high-order correlations of multiple views are ignored.(3)Only special incomplete scenarios such as two-view case can be handled.To address these limitations,this paper mainly proposes two methods as follows:(1)To mine the cross-view relations among data points,explore the complementary information of original data representation,and handle arbitrary incomplete scenarios or data with negative entries,a consensus learning based incomplete multi-view clustering method is proposed.This method combines matrix factorization with index matrix to learn a low-dimensional consensus representation which can exploit the complementary multi-view information from the original feature representation of available instances.In addition,a consensus similarity graph is learned by integrating index matrices and centring term into self-representation learning,aiming to discover the underlying cross-view relations among data points.The key of this method is leveraging graph Laplacian to correlate the consensus representation with the similarity graph,which can reciprocally promote the compactness of the consensus representation and the accuracy of the similarity graph.Extensive experimnets conducted on several incomplete multi-view datasets demonstrates the superiority of the propoed method in IMC task over state-of-the-art methods.(2)Most existing IMC methods cannot exploit the view-specific and cross-view relations among data points and capture the high-order correlations of different views simultaneously.To address this issue,an low-rank tensor bsed incomplete multi-view subspace clustering method is proposed.This method introduces low-rank tensor constraint and centring term into subspace representation learning,aiming to jointly mine the view-specific and cross-view relations among data points,capture the high-order correlations and reduce the redundancies of multiple views.In addition,a novel module is devised to obtain a discriminative similarity graph by approximating the inner product between the view-specific and cross-view subspace representations.This method develops an optimization algorithm based on Augmented Lagrangian Alternative Direction Minimization(AL-ADM)strategy for solving the proposed objective function.The empirical results on several incomplete multi-view datasets validate the effectiveness and advantages of the proposed method.In summary,this paper proposes two novel methods with encouraging performance and flexible applications to address the issues from which most existing IMC methods suffer,provides corresponding optimization algorithms,analyses their computational complexities,parameter sensitivities,and convergence properties,and conducts experiments on several datasets which verifies their effectiveness and superiority.