Researches On Diversity Multi-view Clustering

Many of the data in real life are depicted from multiple perspectives.For example,some news is reported from various sources,and exploring useful knowledge from these data is a major topic of data mining.Most multi-view data is untagged data,and mining useful information from large amounts of unlabeled data is an inevitable challenge.The KMeans clustering algorithm has been proposed for extracting knowledge from large-scale unlabeled data and has been widely used.However,the disadvantage of the traditional K-Means algorithm is that it is only applicable to single-view data.In recent years,many scholars have proposed multi-view clustering methods to overcome the limitation that traditional clustering algorithms can only cluster a single view.However,almost none of these methods considered the correlation of information among views,therefore the high complementarity and low redundancy for each view information participating in the clustering were missing.In this paper,we discuss the existing researches of multi-view clustering from two aspects,then we discuss on the problems of existing multi-view clustering methods.The contributions of our work as follows:(1)Aiming at the insufficiency of existing multi-view clustering methods to balance the consistency and complementarity of information between multiple views effectively,this paper proposes a multi-view KMeans clustering method based on Bregman divergence.In this method,we use a new consistency metric i.e.Bregman divergence to measure the information distortion between clustering results and multiple views.The clustering results are obtained by minimizing the information deviation among the consistency representation and multiple views.In addition,we measure the similarity between each pair of views through the HilbertSchmidt independent criterion,and we measure the importance of each view through the weight vector to ensure that the various view information participating in the cluster has the best complementarity and minimum redundancy.(2)The existing multi-view clustering methods which based on nonnegative matrix factorization(NMF)usually learn a common representation matrix from multi-view data,and then the clustering is performed on the representation matrix by K-Means clustering algorithm.Although existing methods based on non-negative matrix factorization have been widely used due to they are easily understood and their low complexity.However,the existing methods only consider the correlation between the representation matrix of each view and the learned consistent representation matrix.The correlation between multiple view representations is not considered.Aiming at this problem,this paper propose a new joint non-negative matrix factorization multi-view clustering method based on diversity constraint.In this method,we add the diversity constraint of the coefficient representations matrices from multiple views obtained by NMF into the objective function of joint non-negative matrix factorization.We use Hilbert-Schmidt to achieve a measure of the correlation between multi-view coefficient representations,and we control the diversity of view information participating in the cluster by learned weights.In a word,in view of the consistency,complementarity and redundancy of multi-view information,we consider the shortcomings of existing multi-view clustering methods in information fusion,and we propose two multi-view clustering methods in this paper.The validity and feasibility of the proposed method are verified on the experimental data sets.The research in this paper provides new ideas and new methods for multi-view clustering,and has certain theoretical value and application value in the field of multi-view clustering.