Research On Multi-View Deep Graph Clustering

Mulit-view data are often captured from various sources or composed of multiple modalities,endowing computers with the ability to understand it has become a key issue in data mining.As an unsupervised learning method,clustering analysis is superior in mining multi-view data’s intrinsic structure and latent information and has drawn wide attention from researchers.Recently,to solve the shortcomings of traditional methods in effectively processing high-dimensional data,deep clustering methods use deep neural networks to perform nonlinear transformations on data,then achieve better clustering analysis in a new feature space.However,previous methods still need to improve synchronously learning data representation and clustering processes and understanding heterogeneous data from multiple sources.To solve these issues,this paper proposes a multi-view deep graph clustering method and verifies the proposed model’s effectiveness,robustness,and parameter sensitivity.The main contribution of the paper is summarized as follows:(1)Previous deep clustering methods separate the data representation learning from the clustering process without fully considering the connection between the representation learning and the clustering process,which limits the final clustering performance.Therefore,this paper proposes a deep clustering method based on graph convolutional networks,which uses graph convolutional networks to map high-dimensional data into a low-dimensional embedding space for clustering analysis,and combines the data representation and clustering processes in the same optimization process.In addition,graph neural networks require labels for supervised learning,while previous methods often require labels for semi-supervised learning.In response,this paper designs an adaptive pseudo label generation strategy that assigns pseudo labels to data points in closed-form solutions and optimizes them jointly with deep embeddings,then continuously improving the plausibility of the pseudo labels.Finally,the model is trained by fitting pseudo labels with clustering assignments in an unsupervised learning pattern.The experimental results show that,compared with six classical clustering algorithms and state-of-the-art deep clustering algorithms,the proposed method significantly improves the average clustering accuracy by 3.75% and the average normalized mutual information by 4.18% on four publicly available graph benchmark datasets improving the clustering performance.(2)Previous multi-view clustering methods focus on the topological structure information of data but ignore the feature of original data,which leads to insufficient information to be learned and captured.As a result,the model performance is highly dependent on the constructed similarity graph.To address this problem,this paper proposes a multi-vieworiented deep graph clustering method based on the deep clustering method based on graph convolutional networks.Firstly,the feature learning capability of the model is improved by the proposed deep clustering method based on graph convolutional networks considering both feature information and structural information of the data.Secondly,to address the subproblem that previous graph construction methods require much experience for parameter tuning and poor robustness,a similarity graph construction method is designed,which can adaptively assign optimal neighbors to data points and achieves effective and stable performance on different types of datasets.Finally,consistent representation learning of multiview datasets is performed through an adaptive weighted graph fusion strategy to achieve cross-view structural information fusion.The experimental results show that compared with eight classical clustering algorithms and state-of-the-art multi-view clustering algorithms,the proposed method achieves an average accuracy improvement of 5.78% and an average normalized mutual information improvement of 2.91% on six different types of open multiview datasets;compared with deep clustering methods based on graph convolutional networks,the proposed method achieves a 13.63% improvement in clustering accuracy on multi-view data.The effectiveness and robustness of the proposed multi-view deep graph clustering method for multi-view clustering are well demonstrated.