Incomplete Multi-View Clustering Based On Common Representation Learning

With the continuous increase of data collection methods,real data often consists of multiple modalities or comes from multiple sources,which is called multi-view data.The machine learning task based on multi-view data is called multi-view learning.Nowadays,as a hot field of multi-view learning,multi-view clustering has attracted wide attention of researchers,which provides a way to partition multi-view data into their corresponding clusters.At present,most multi-view clustering algorithms assume that all views are complete.However,in the real-world applications,it is common that each view may contain some missing data instances,resulting in incomplete multi-view data.The existing multi-view clustering algorithms cannot be directly applied to incomplete multi-view data,and clustering such data is called incomplete multi-view clustering.To solve the problem of incomplete multi-view clustering,the research idea of this paper is as follows: It is necessary to learn a common representation of all views for clustering.When learning the common representation,an error matrix is introduced to model the missing instances.In addition,the affinity matrix is learned to preserve the global structure of the data,so as to learn a more discriminative and compact common representation.Finally,the adaptive weight learning is introduced to learn a more appropriate weight for each view to measure the importance of the view,and at the same time,a more discriminative common representation is obtained,so as to improve the clustering performance.This paper takes incomplete multi-view clustering as the research object,and proposes the following research results:(1)Simultaneously Learning the Common Representation and Affinity Matrix is proposed.This method learns a latent common representation for all views for clustering.At the same time,it takes into account exploiting the available information of the non-missing views and the underlying semantic information of the missing views to learn common representation.In addition,this paper introduces an error matrix to model the missing instances.By recovering the missing views,all incomplete views are naturally aligned to learn the latent common representation of the views.In order to preserve the global structure of the data,this paper proposes to learn the affinity matrix of the data while learning the common representation.Simultaneously learning common representation and affinity matrix can guide to learn more discriminative and compact common representation,thereby improving clustering performance.Finally,comparative experiments are carried out on four real datasets,and the experimental results verified the effectiveness of the algorithm.(2)Adaptive Weighted Incomplete Multi-View Clustering is proposed.This method takes into account that due to the number of available instances and feature dimensions of each view,the available discriminant information of each view has a large difference,which leads to the situation that the importance of each view is different.Due to the lack of prior knowledge,it is impossible to assign each view an appropriate weight in advance to measure the importance of the view.In order to measure the importance of views and improve clustering performance,this paper adds adaptive weight learning to the model of the common representation and affinity matrix learning.By learning the weight of each view,the most reasonable weight of each view can be obtained,and at the same time,assigning appropriate weight to each view can help to learn a more discriminative common representation.Finally,experiments are conducted on five real datasets,and the experimental results verify the effectiveness of the algorithm.