Research On Clustering Methods Via Structure Representation Learning

Clustering is a basic tool of data mining,which is widely used in the fields of computer vision and pattern recognition.The core content of clustering is to divide samples by comparing the similarity of data without prior information of categories.The similarity of samples has a good correspondence with the spatial structure of samples.Therefore,mining the structural information of samples is an important part of clustering.Subspace clustering is a typical method of clustering by using the structure information of high-dimensional data in subspace.With the help of representation learning and structure regularization in subspaces,many subspace clustering methods have achieved remarkable results.At present,most of the subspace clustering methods study the structure information of the whole data,but seldom analyze the specific structure of the data.Due to the strong representation capability of the deep network,clustering based on deep representation learning has become an important research direction.Most existing deep clustering methods combine unsupervised representation learning and class division to design deep clustering models.Some deep clustering methods only explore the information of the data itself to obtain the representation in the hidden space,and seldom consider the structural information between data samples.Other studies mainly study the structural information between data while ignoring the characteristics of the data itself.Therefore,the accuracy of clustering is limited.To solve the above problems,two improved methods are proposed in this paper:Firstly,to address the problems of the subspace clustering methods in analyzing the concrete structure of data,and to deal with the relevant problems in the sequential data clustering,a clustering method based on subspace support structure representation learning is proposed.In the method,The initialized adjacency matrix is used to establish the corresponding spatial structure relation graph,then the graph is used to learn the support structure representation of data.Under the premise of ensuring both the sparse connection relationship between samples,the inherent structure of subspace,and the complete structure of the cluster,the proposed method can learn the subspace representation of data with structural information,and then the representation is used to compute the similarity matrix and clustering.Secondly,by integrating the basic representation of the auto-encoder network into the graph convolution network,a clustering method based on the deep graph convolution structure representation learning is proposed.The proposed method consists of three main parts: the basic module uses the graph clustering loss and reconstruction loss to ensure the basic learning ability of the network;The structural representation learning module based on graph convolution uses the self-supervised Kullback-Leibler divergence constraint to obtain the representation of structural information with data.The contrastive learning module uses contrastive loss to enhance the discriminability of learned representation.Experimental results show that the proposed method can learn the structure information of data by utilizing the structure regularization method based on graph convolution,so as to improve the clustering results on the dataset.The experimental results and theoretical analysis show that the method based on subspace support structure representation learning can improve the clustering results on the basis of ensuring the complete structure of clusters and the inherent structure of the subspace.And the method based on deep graph convolution structure representation learning can effectively improve the clustering performance.