Community Detection Based On Attributed Network Representation Learning

Graph or network is a common data structure for exploring and modeling complex systems in the real world,such as social networks,citation networks and co-authorship networks.The community structure widely exists in the real complex network.Finding community structure in networks is of great significance to the study of complex networks.With the increasing scale of network data,nodes in the network are usually associated with some attributes.The traditional community detection model only captures the shallow connection relationship,while the attribute network representation learning can capture the network topology and node attribute information,and learn to get a unified vector representation.It has become a new research task to detect community structures in complex networks by keeping network structure information and node attribute information at the same time.Therefore,this thesis improves the community detection method based on attribute network representation learning focusing on non-overlapping community detection and overlapping community detection,so as to improve the quality of community division.The main work and innovations include the following parts:(1)A non-overlapping community detection model fusing graph attention network is proposed.Firstly,the auto-encoder is used to encode the node attribute features and the graph attention network is used to capture the network structure features.The graph attention network calculates the influence weight of the neighborhood node on the target node through the attention mechanism,taking into account the difference of the influence of different neighborhood nodes.At the same time,high-order neighborhood information is added to the graph attention layer.And the vector representation of the node space position is obtained.The representation of each layer learned by the auto-encoder is transferred to the graph attention layer by balancing parameters,which realized the fusion of structure information and attribute information.Finally,the dual self-monitoring mechanism is used to update the whole model,and the end-to-end clustering task is realized.Through the experiments on six datasets,the results show that the model performs well in most datasets,especially for the networks with longer average path length and network diameter and smaller clustering coefficient.The value of metrics on each dataset can increase up to 4.3% on ACC,2.8% on NMI,4.2% on ARI and4.6% on F1.(2)A non-overlapping community detection model based on feature fusion is proposed.In the first part,the first fusion of structure and attribute features is realized by balancing parameters.But there will be information loss in the process of transferring the features along with the neural network,which leads to the lack of full fusion of structure and attribute features.Therefore,on the basis of the first part,this thesis designs a feature fusion module.Firstly,the hidden layer representation of attention layer is extracted for non-local operation,then the node representation obtained by the first fusion is multiplied with it,and the structure and attribute features are fused for the second time.The experimental results show that the feature fusion module can extract the most relevant features of clustering tasks and further improve the effectiveness of community division.The value of metrics on each dataset can be improved by up to 2.4% on ACC,3.6% on NMI,4.3% on ARI,and 2.0% on F1.(3)An overlapping community detection algorithm based on Bernoulli-Poisson model is proposed.Based on the assumption that the edges between the overlapping nodes in the community are dense,the model adds node clustering coefficients to capture the close connection between nodes on the basis of the graph attention network integrating the high-order neighborhood information of nodes.Then the network structure and node attributes are used as the input of the graph attention network.The negative logarithmic likelihood function of Bernoulli-Poisson model is used to optimize the model parameters to learn the node-community membership matrix.And then the matrix is mapped to obtain the overlapping community.Experiments are carried out on five datasets,and the overlapping NMI is used as the metric.The experimental results show that the model performs better on most datasets.