Research On Multiview Clustering Method For Imperfect Data

Clustering has been widely used in many machine learning tasks in recent years.Traditional clustering methods usually use a single view to measure the similarity of samples.However,with the rapid development of data collection,a single feature is no longer sufficient to fully describe a data sample.Multiview data can provide more information in revealing the underlying clustering structure than the traditional single-feature representation.However,in real life,due to various objective factors,phenomena such as missing data and noise interference often occur,and such data are called imperfect data.The existence of imperfect data can lead to the degradation of the performance of existing multiview clustering methods.To face this problem,this paper proposes three multiview clustering methods for imperfect data.The specific research contains the following aspects.(1)A multiview clustering method based on low-redundancy representation learning is proposed.The method exposes the redundancy of high-dimensional data in the kernel space,learns the low-redundancy data representation of local views in the kernel space,and uses it for subspace learning.The obtained subspace representation matrix is then stacked into a third-order subspace tensor,which is constrained using the tensor kernel norm.The similarity information within views and the complementarity information between views are fully exploited.Extensive experimental results show that the proposed method outperforms current mainstream multiview clustering methods in several objective evaluation metrics and is more suitable for multiview clustering of high-dimensional data.(2)An incomplete multiview clustering method based on a weighted low-rank tensor is proposed.Firstly,in order to learn the complete subspace representation,it is proposed to implement the missing data in the kernel space with multiview complementation.And the preservation of local information of views and the fusion of global information are fully considered.Then,a weighted low-rank tensor kernel norm is introduced to generate the optimal common representation matrix according to the importance of different views.A large number of multiview clustering experiments with different missing rates on three publicly available datasets show that the performance of the method is advantageous in most cases compared to the current mainstream incomplete multiview clustering methods.(3)An incomplete multiview clustering method based on a joint complementary kernel representation is proposed.On one hand,the method merges kernel compensation model and feature representation learning into a unified framework where the two processes can be negotiated with each other.On the other hand,the idealized representation of a tensor constructed from multiple similar matrices is found using tailored low-rank tensor constraints,making full use of inter-view and intra-view information.Experimental results on four benchmark datasets show that the proposed method outperforms seven state-of-the-art incomplete multiview clustering methods in three metrics: accuracy,normalized mutual information,and Fscore.The performance of incomplete multiview clustering is improved.