Research On Deep Multi-view Joint Clustering

With the coming of big data era,the data may be collected from different sources.More and more feature extractors have been designed to obtain distinct features of data as well.As a result,multi-view learning has raised wise research interest among scholars.Multi-view clustering is an important research topic in the clustering area.Exploiting the complementary information among multiple views sufficiently can enhance the clustering performance.This thesis focuses on multi-view clustering.In order to alleviate the separation characteristic of existing multi-,view clustering methods,this thesis develops deep multi-view clustering algorithms based on joint learning,which can improve the performance of multi-view clustering effectively.The main contributions of this thesis are summarized as follows:Firstly,a deep multi-view joint clustering model based on explicit multi-view fusion is proposed.Since different views have different importance levels in clustering,novel explicit multi-view weights are imposed on different views.With the construction of multi-view fusion auxiliary target distribution and a soft assignment distribution,the model can realize the joint learning of multi-view features,clustering assignments and explicit multi-view weights.The model is optimized via a KL divergence like clustering objective and an additional regularization.Experimental results on hand-written digital datasets,object datasets and scene datasets demonstrate that the proposed model is superior to existing two-step multi-view clustering methods and single-view joint clustering algorithms.Secondly,a deep multi-view joint clustering model based on implicit multi-view fusion is proposed.Since clustering centroids in different views are related,novel implicit multi-view weights are imposed on them.A novel multi-view fusion soft assignment distribution and an auxiliary target distribution are constructed,which can support the joint learning of multi-view features,clustering assignments and implicit multi-view weights.The model is optimized via a KL divergence like clustering objective.The proposed model is evaluated on hand-written digital datasets,object datasets and scene datasets,and the results show the superiority of the proposed model over other comparing methods.Thirdly,a deep multi-view joint clustering model based on self-paced multi-view fusion is proposed.Since different views and samples have different complexities,a novel self-paced learning scheme is introduced in the proposed model,and multi-view sample weights are imposed on different samples in each view.The proposed model can realize the joiint learning of multi-view features,clustering assignments and multi-view sample weights.The model is optimized via a multi-view K-means clustering objective and an additional autoencoder reconstruction term.Experiments are conducted on hand-written digital datasets and object datasets,and the results show the feasibility and effectiveness of the proposed model.