Research On The Construction Method Of Group Profiles Combining Network Structure And Text Content

The rapid development of information networks has led to an increase in user data.Research on the construction of user profiles based on network data is of great significance for research field of personalized services such as accurate recommendation.Aiming at the problems of low user modeling accuracy and poor network group similarity and tightness caused by the existing group profiles construction methods,which are mostly based on the text content published by users on the network and less on the network structure information,a group profiles construction method combining network structure and text content is proposed.First,the local and global network structure is modeled using the first-order proximity and second-order proximity of the LINE model,and the local and global structure modeling vectors are combined through objective function optimization.With the attention mechanism,an attention matrix is introduced into the hidden layer of the neural network to represent the context of the text content,and then the text content modeling is realized.The two models are combined and trained by convolutional neural network to represent network users as spatial vectors.Secondly,the density peak clustering algorithm is used to cluster the user space vectors.By calculating the local density and distance of each user,the user category label is determined,and the clustering result is iteratively optimized using the structure-content modularity to achieve the construction of network groups.Finally,the LDA topic modeling method is used to divide the topics to obtain the topic tag of each group,and the visual tools are used to depict the group profile.Three data sets,zhihu,Cora and Hep Th,are used to compare the accuracy of modeling and the effect of group construction.Compared with six network user modeling methods such as Deep Walk,LINE,Node2 vec,SDNE and TADW,the precision@k in this paper all achieves the optimal value,and the MAP is improved by 0.1,0.12 and 0.14 respectively,and the AUC is improved by0.02,0.04,and 0.03 on average.Compared with three group construction methods such as Louvain,SA-Cluster and based on K-means clustering,the density value of the method in this paper increases by 0.27,0.073,0.05 on average,indicating good intergroup tightness.Compared with the two structure-based methods of Louvain and sa-cluster,the entropy value decreases by0.35 and 0.46 on average,indicating a high similarity among the groups.There are 23 graphs,14 tables and 57 references in this paper.