Research On Multi-view Clustering For Complex Mapping Relationship And Noisy Data

In the past few years,many outstanding multi-view clustering algorithms have been proposed.Most all of them,whether it is for complete multi-view data or partial multiview data,only consider that the samples between different views are strictly one-to-one relationship or no mapping relationship.They ignore the complex relationship between cross views data samples,in which a sample in one view may correspond to multiple samples in another view in varied degrees.On the other hand,it is common that a view may be mixed with some noise data,which affects the clustering effect.Therefore,this thesis studies the complex relationsip between cross views data and the problem of noise in multi-view data.To solve the problem of complex mapping relationship between multi-view data,this thesis proposes a novel multi-view clustering method,named complex mapping multi-view clustering,which obtains high quality clustering results by mining and using the complex mapping relationship between views.Specifically,the method firstly constructs a complex mapping relationship matrix for each pair of views by using the nearest neighbor relationship between the samples within view.Then the complex mapping relationship matrix applied to the framework of multi-view clustering based on non-negative matrix factorization to guide multi-view information fusion in order to obtain more compact representation of multi-view data space.Finally,an effective optimization scheme is provided for the proposed method.Aiming at the problem that multi-view data contains noise,this thesis proposes a novel multi-view three-way clustering method for noisy data.With the idea of matrix decomposition,the method decomposes the similarity matrix of each view into the good data and the corruptions,and uses only the good data for the next process multi-view information fusion.Since the existing multi-view clustering methods are mostly hard clustering methods,and it is difficult to describe the uncertainty relationship between cluster and objects.Therefore,this thesis utilizes the three-way clustering representation to obtain the final uncertainty clustering results.Finally,we design corresponding comparative experiments to verify the effectiveness of the two methods proposed in this thesis.The experimental results show that by using the complex relationship between views,better experimental results can be obtained both in the complete multi-view data and partial multi-view data.In processing multi-view data contaminated by noise,our method can obtain better clustering results than the comparison method.