| With the vigorous development of the Internet,the National Fourteenth Five-Year Plan proposes to vigorously develop the digital economy,and network data has exploded.Graphs are widely used in the real world to represent complex relationships between entities,such as citation networks,transportation networks,and social networks.Considering the manpower and material resources required for massive data labeling,from traditional statistical inference to recent deep learning methods,graph clustering tasks have been widely studied.As a basic task in the field of data mining,attributed graph clustering aims to divide the nodes of the network into unconnected clusters.How to effectively integrate the two heterogeneous information in the graph,namely network topology and attributes,is an important and difficult point of attributed graph clustering and graph representation learning.The graph convolutional network effectively combines the network topology and node attributes through node aggregation and feature transformation,and is the main tool for current graph learning.As an unsupervised task,attribute graph clustering naturally lacks label information.Therefore,general graph representation learning methods cannot guarantee that the learned representation is suitable for subsequent clustering tasks,and self-supervised learning can effectively add task-related constraints.Based on this,this paper proposes a dual channel graph convolutional network model.Through self-supervision,the low-dimensional representation of constrained nodes approximates the clustering center of the soft target distribution,and the representation related to the clustering task is learned.The loss function of dual channel graph representation learning usually adopts topology reconstruction and attribute reconstruction respectively,while the low-dimensional representation learned based on attribute reconstruction only reflects the high-level semantic information of the attribute,and does not contain the topology information that characterizes the relationship between nodes.Therefore,it is unreasonable to directly restrict the consistency of the two channel representations.Based on this,this paper proposes a topology-preserving multi-channel fusion graph convolutional network,adding structural constraints to the low-dimensional representation of the attribute reconstruction channel,and then fusing the two features,using a self-supervising mechanism to constrain the three channels to represent the topological consistency.Current graph learning is generally based on an assumption: the nodes connected by edges on the graph tend to belong to the same class.We use homology to represent the proportion of nodes connected by edges belonging to the same class.In the real world,there are often data with low homology.For example,the principle of "opposite sex attracts" on dating sites,users are more inclined to communicate with the opposite sex.In this case,it is not appropriate to directly use the graph convolutional network to make a graph to indicate that the learning is not appropriate,and it will aggregate the features that do not belong to this class through the noise edge.We propose a graph convolution method for topology masking,which adaptively removes edges in the topology that may be noisy,then we designed a topological mask graph autoencoder with a symmetrical structure in this paper.This paper compares with the current state-of-the-art algorithms on multiple data sets and proves the effectiveness of the three algorithms mentioned above. |
