Research And Implementation Of Incomplete Multi-View Clustering Algorithm

With the rapid development of information collection technology,data is usually collected from different sources or described by different feature subsets,which make up multi-view data.Multi-view data contains more abundant information than single-view data,which can effectively improve the performance of downstream learning tasks,so multi-view learning has become a research hotspot in the field of machine learning and data mining.Intuitively,the increase in the number of views causes a sharp increase in the labor cost required for sample labeling.Therefore,multi-view clustering,a fundamental unsupervised learning paradigm,has received extensive attention.Existing multi-view clustering algorithms usually assume that samples have complete representations in multiple views,however,in practical applications,incomplete multiview data is ubiquitous.In order to effectively divide incomplete multi view data points,incomplete multi-view clustering has been significantly developed in recent years,but there still exist some key issues to be further explored: how to determine the weight of different views in the view fusion process;how to unify feature learning and clustering to avoid suboptimal results;how to fully explore non-linear relationships among multiple views.To solve the above problems,this paper completes two innovative works around incomplete multi-view clustering.An incomplete multi-view clustering algorithm based on linear neighborhood reconstruction is proposed,which learns the multi-view latent representation through weighted matrix decomposition mechanism.Specifically,in the process of matrix decomposition,it is assumed that each view has its own basis matrix and shares the multiview unified representation,which effectively solves the view weight assignment problem and can easily handle data with more than two views Then,we use the linear neighborhood reconstruction technology to reconstruct the latent representation matrix.The reconstruction relationship between samples is explored in the neighborhood of the samples.The obtained self-representation matrix exhibits a clear block-diagonal structure and correctly reveals the essential structure of multi view data.Finally,a similarity matrix is calculated based on the self-representative matrix and spectral clustering is applied to obtain the final partition result.In conclusion,this proposed model combines latent representation learning,self-representation learning,and spectral clustering into a unified framework,and optimizes the solution in a unified manner to ensure that the learned feature representation is optimal.Extensive comparative experiments show that the proposed model has obvious advantages over the existing clustering algorithms.The self-supervision based deep incomplete multi-view clustering algorithm is proposed,which utilizes a set of view-specific auto-encoders to learn a low-dimensional representation for each view,and the auto-encoders endow the model with nonlinear capabilities.Then,through two different feature fusion mechanisms: weighted average and feature concatenation,different levels of feature fusion are considered.After obtaining two common representations,two label prediction matrices are obtained through a fully connected classification network,and the consistency among multiple views is mined by maximizing their mutual information.Two groups of auto-encoders are connected between the two label prediction matrices,which are each other’s input and output.In this way,this method realizes self-supervision mechanism and provides supervision information for feature learning.In the testing stage,the final clustering result can be directly obtained by the weighted average of two label prediction matrices.Finally,on four real data sets,the experimental results compared with seven benchmark algorithms verify the effectiveness and comparability of this algorithm.