Social Network Link Prediction Based On Node Feature Similarity

In recent years,the development of Internet technology has produced a large number of online social activities,coming with a large number of social network data.The analysis of social network data is of great significance in the analysis of public opinion in information network,the prediction of users' preferences in tourism system and the recommendation of friend relationship on social platforms.As an important method of social network analysis,link prediction has great academic value and application value.Although many approaches from fields and disciplines have been proposed,there are still some shortcomings in the current social network link prediction methods.First,the simple attribute information extraction method is not accurate enough.Secondly,the performance of link prediction methods using only node attribute information or first-order adjacency relationship is greatly affected by network sparsity.Thirdly,many link prediction methods can only predict the existing links,but few people make clear calculation of the non-existent links.In view of the above problems,this paper mainly does the following works:(1)Aiming at solving the problem that the method of using only single similarity information in attributed networks is greatly affected by network sparsity,and that the extraction of network information is not accurate enough,a link prediction method based on combine similarity and indirect transportation of similarity(CS&STrans)is proposed to predict the cooperative relationship in scientific research networks.When extracting the text attributes of nodes,the influence weights of different original texts on the current nodes are taken into account;the first-order similarity and attribute similarity of nodes are allocated dynamically;in addition,the second-order similarity of nodes is mined via the indirect transportation mechanism of node similarity,and the comprehensive similarity is obtained.The experimental results show that the performance of CS&STrans method is better than that of single method in the process and other benchmarking methods.At the same time,parameter comparison experiments also show that the variable parameter settings of the method make the algorithm have better scalability and can be applied to different attribute networks.It has great application value in the actual attribute network analyzing system.(2)In order to predict the link that does not exist in the test data and improve the stability and convergence speed of the parameter model in the process of node vectorization,a link prediction method based on node mapping and label data construction(Net2Vec-CLP)is proposed.First,in the process of node vectorization,the Newton method is used to optimize the model parameters and the auxiliary function is introduced to avoid the problem that the gradient of the optimization process may be zero.Finally,labeled data sets are constructed for classification learning,and the classification results of node vectors are obtained as link prediction results.The validity and feasibility of the proposed method are analyzed on multiple data sets and multiple evaluation indicators.The experimental results show that Net2Vec-CLP has a high prediction accuracy,and has obvious advantages over the benchmark method.The high accuracy results of the algorithm make it have better potential application value in friend recommendation,preference analysis and interest prediction in social network systems.