Clustering Analysis On Multi-view Matrix Factorization

With the rapid development of multi-media technology,more and more information can be obtained,which includes a large number of useless information and much invisible information.Therefore,data mining is an very important tool to analyze information in machine learning and AI fields.Clustering analysis is unsupervised so that it can cluster datasets without any prior knowledge,making the samples as similar as possible within clusters and as different as possible among clusters.Therefore,as an important method of data mining,clustering analysis is of great significance in the fields of pattern recognition and artificial intelligence for its abilities to reveal the relationship among the samples.Multi-view clustering analysis can describe the database more comprehensively because of its use of multiple views,thus it can perform better than single-view clustering.The traditional multi-view clustering algorithms only consider the consistency among different views,neglecting the similarity among the different samples in the same view before and after the feature learning.Therefore,this paper proposes a multi-view clustering algorithm called MLN by introducing local spatial structure constraints into the traditional multi-view learning framework.However,the MLN algorithm is based on NMF,so it is only applicable to nonnegative matrices.In real worlds,data is negative sometimes.Therefore,this paper proposes a multi-view clustering algorithm called MLSN based on SemiNMF.MLSN performs well whether the data matrix is nonnegative or not.Experiments on three datasets show that MLN and MLSN have good clustering performance.As an ensemble clustering method,evidence accumulation clustering(EAC)can integrate a group of different base clusterings to improve the generalization ability of clustering analysis and obtain higher clustering accuracy.EAC is divided into two steps: constructing co-association matrix and performing hierarchical agglomerative clustering.However,the traditional weighted EAC only takes into account the influence of the base clusterings,or only considers the influence of cluster qualities.In fact,the two will affect the correlation among the samples.Therefore,this paper proposes a new weighted EAC method called GLWEA.When the co-association matrix is constructed,the quality of the base clusterings and all of the clusters will be taken into account,and then the final clustering results is obtained by hierarchical agglomerative clustering algorithms.The experimental results show the effectiveness of GLWEA.When multi-view clustering is combined with ensemble clustering,the influences by different parameters can be reduced to improve clustering performances.Therefore,this paper pro-poses two ensemble multi-view clustering methods,called LEMGSN and GLEMGSN respectively,by combining two EAC methods LWEA and GLWEA with MLSN.Experimental results show that LEMGSN and GLEMGSN are better than MLSN.