View-Consistent Self-Supervised Graph Clustering

Graph clustering algorithms are widely used in real-world problems such as data mining and community detection.The core step of graph clustering is to extract feature embeddings suitable for clustering from original feature data.In recent years,deep graph neural networks have developed rapidly.However,they all have their own limitations and cannot extract sufficiently deep and comprehensive feature embeddings.In addition,how to combine graph-structured features and attribute features for clustering is also a problem to be considered.The main contents of this paper include the following three points:(1)We propose two View-Consistent Self-supervised Graph Clustering(VCSGC)algorithms.Specifically,the original attribute feature matrix is mapped into two-view feature embeddings,and then the deep graph convolutional neural networks are used for feature extraction of the respective views.Finally,the consistency loss function is used to obtain the view-consistent feature embeddings for clustering.By utilizing two different deep graph neural networks to extract multi-view feature embeddings,we finally obtain more comprehensive feature embeddings for clustering.In addition,a view consistency loss function is designed in this paper and its validity is verified.(2)We propose a triple self-supervised module,which extracts more comprehensive feature embeddings for clustering.Firstly,the pre-trained autoencoder network is used to extract the attribute feature embedding,and then the k-means clustering algorithm is used to generate the soft label.Finally,the sharpened soft label is generated by taking squares and regularization,which is used to supervise the learning of the autoencoder and the deep graph neural networks of the two views.This module unifies the attribute feature embeddings extracted by the autoencoder and the graph-structured feature embeddings extracted by the deep graph neural networks into one learning objective,so as to extract more comprehensive feature embeddings for clustering.(3)The proposed two algorithms are tested on three single-view graph-structured datasets.The advancement of the algorithms proposed in this paper is verified by comparing them with other deep graph clustering algorithms.In addition,ablation experiments and model-robustness experiments are carried out to verify the effectiveness of the proposed algorithms and the robustness of the model performance to hyperparameters.This paper first briefly describes the research developments of clustering algorithms,graph clustering algorithms and multi-view learning algorithms,and then introduces several graph clustering algorithms in detail.Next,the two View-Consistent Selfsupervised Graph Clustering(VCSGC)algorithms proposed in this paper are described in detail.Finally,experiments are carried out on three single-view graph-structured datasets,and the experimental results are displayed and discussed.