Research And Application Of Community-enhanced Attributed Network Representation Learning

With the rise of large-scale social networks,network mining has become an important sub-domain of data mining.In network mining,the feature representation of nodes is very important.The traditional approach of nodes' representation based on adjacency matrix brings severe challenges to the processing and analysis of large-scale networks due to the problems of high dimension and sparsity.It is of great significance to find low-dimensional effective representation of network for solving practical application problems in real networks.Network representation learning aims to map the nodes in the network into a low-dimensional,continuous real-valued semantic space so that the nodes with large similarity have similar vector representations.Network representation learning can effectively alleviate the problem of sparse data of large-scale network,fully integrate heterogeneous information of network,and improve the efficiency of large-scale network analysis and processing.As the most widely studied and applied algorithm in network representation learning,attributed network representation learning integrates network structure information and node attribute information,which can fully explore the latent semantic information of network nodes,and has become one of the current research hotspots in network representation learning.By early fusion of network structure information and node attribute information,the attributed network representation learning method SNE can effectively play the role of mutual supplement and restriction between network structure information and node attribute information,and obtain better node embedding representation.This paper focuses on SNE,and the main research contents are as follows:(1)In view of SNE's shortcoming that it can not well maintain the global structure of the network,it is expanded the model by adding clustering constraint information,forced nodes to carry out k-means clustering to reveal the community structure in the network,so that the learned node vector representation can better maintain the global potential clustering structure of the network.Finally,clustering and classification experiments on five data sets indicates that the improved method could maintain the global structure of the network better.(2)In the real attributed network,there is also a small amount of node category label information in addition to network topology and node attribute information.The introduction of the node's category label information enables the learning of the node embedding representation to be discriminative.Therefore,on the basis of the previous research,it is continued to consider the small number of prior node category label information to further enhance the embedded community information of network nodes.Finally,experiments are taken on four real data sets,and the experimental results show that the node embedding representation learned by the improved method contains more community structure information.(3)The feature vectors of CSDN users are extracted by using the attributed network representation learning method,and the users' interest is marked.Experimental results verify the effectiveness of the attributed network representation learning method in user feature extraction.