Research On The Multi-view Clustering Algorithms Based On NMF

Without some powerful means of data mining,it is too difficult for people to compre-hend data to get information.The growing gap of data and information gives birth to the technology of data mining.Clustering is a strong tool to dig the data,which clusters instanc-es to make the instances in one cluster more similar and the instances not in a cluster more different.Multi-view data coming from multiple ways or being presented in multiple forms,has more information than single-view data.Previously,the conventional clustering algo-rithms are applied to deal with the multi-view data.They merge the multiple views into one single view leading to over-fitting.Moreover,if the size of data is very small and each view has its own statistical property,even there is no physical meaning in the process of clustering.Thus multi-view clustering appeared.Nonnegative matrix factorization(NMF)is an effective method to learn low-rank ap-proximation of nonnegative matrix of nonnegative data,but it may not expert in clustering.This paper proposes two multi-view clustering algorithms and a clustering ensemble algo-rithm based on NMF.This paper presents a novel multi-view clustering algorithm,which properly combines the similarity and NMF(called MVCS).It aims to obtain latent features shared by multiple views with factorizations,which is a common factor matrix attained from the views and the corporate similarity matrix.Besides,according to the reconstruction precisions of data ma-trices,MVCS could adaptively learn the weight.In addition,we propose a novel multi-view clustering algorithm via joint nonnegative matrix factorization to integrate the two latent fea-tures from multiple views(called LFNMF).The key point is to formulate a joint nonnegative matrix factorization process under the constraint of factor similarity matrix among different samples to explore the shared latent feature.For attaining the complementary feature,we set up a basic similarity matrix about the dimensional information.Then the base space of all the views is uniformed to maximize the complementary space each view has.Our multiplica-tive-based algorithm of LFNMF is deduced and proved convergence.The method is com-pared with eight baseline multi-view clustering algorithms in six data sets.All experimental results on our method show superior clustering performances.A NMF-based K-means clustering ensemble(called NBKCE)is proposed for solving the problem of effective information loss in ensemble,which is caused by basic clustering results obtained from the original datasets.In NBKCE,an ensemble information matrix is built by exploiting the results of the K-means primarily,and then the relationship matrix is formed based on the original dataset.At last nonnegative matrix factorization(NMF)is em-ployed to construct consensus function to gain the final results.The experiments demonstrate that the NBKCE may attain the underlying information effectively and improve the perfor-mance of the clustering.