Incomplete Multi-View Clustering Analysis Based On Deep Learning

With the rapid development of the Internet era,different types of data can be easily accessed and clustering analysis using multi-view data has become a popular research area.The advantage of multi-view clustering is that it can effectively divide things into different categories due to the richer information contained in multi-view data.This approach is widely used in areas such as search engines,intelligent recommendation and artificial intelligence.However,due to multiple unknown factors,some views of multi-view data usually have missing data.This makes incomplete multi-view clustering analysis draws more and more attention.Existing incomplete multi-view clustering methods suffer from two problems: first,they focus too much on estimating or recovering the missing data and ignore the fact that the estimated values may be inaccurate due to unknown label information.Second,the common features of multi-views are always learned from the complete data,while ignoring the differences in feature distributions between complete and incomplete data.Therefore,two approaches are proposed in this paper to solve the above problems:(1)Adaptive feature projection with distribution alignment for deep incomplete multiview clusteringFor the noise problem caused by data prediction filling,this paper proposes an interpolation free depth incomplete clustering method.In this paper,we propose an interpolationfree depth incomplete clustering method,and consider the distribution alignment in feature learning.The method is a depth incomplete clustering without interpolation.An auto-encoder is used to learn the features of each view and adaptive feature projection is used to avoid the interpolation of missing data.All available data are projected into a common feature space,and the common clustering information is explored by maximizing mutual information,and distribution alignment is achieved using minimization of mean differences.Extensive implementation results show that the proposed APADC method achieves superior results on four benchmark datasets compared to recent state-of-the-art methods.(2)Incomplete multi-view clustering based on intra-cluster distribution alignmentTo address the problem that global feature alignment may cause clustering information omission,this paper proposes a multi-view clustering analysis from the perspective of cluster classes,aligning cluster class distribution to achieve more fine-grained information exploration,so as to obtain the structural information that is beneficial to clustering.The method designs a formula for calculating the weights of the samples by constructing the soft-label features of the view samples,and uses the local maximized mean difference metric in sub-domain adaptation to achieve the alignment of cluster class distribution,while aligning the global features of the views.Extensive experimental results show that the IMVC-ID method achieves better results than not only the recent similar methods but also the APADC method proposed in this paper on four clustering common datasets.In this paper,we propose the depth incomplete multi-view clustering method based on distribution alignment adaptation and the cluster class distribution-based depth incomplete multi-view clustering method.In this paper,we propose the depth incomplete multiview clustering method based on distribution alignment and the cluster class distribution based depth incomplete multi-view clustering method to solve the problem that multi-view clustering cannot be achieved due to missing views.The effectiveness and feasibility of the proposed two methods are verified on some benchmark datasets commonly used for clustering tasks.