A Study About Link Prediction In Social Networks Combining Social Influence And Homogeneity

Network data mining has numerous applications in many disciplines including telecommunication networks,transportation networks,and social networks.The explosion of network data has not only created new opportunities,but also new challenges.Among numerous research problems related to network mining,link prediction is of fundamental importance in numerous tasks in social networks.The problem of predicting relationships in networks is called link prediction.Link prediction aims to infer the behavior of the network link formation process by predicting missed or future relationships based on currently observed connections.In social networks,the establishment of social links is not only determined by personal intrinsic interests but also by neighbors' influences in interpersonal relationship.The influence of neighbors may vary across different neighbors.However,the independent influence of each neighbor has not been separately considered in current link prediction approaches.Furthermore,influence of each neighbor may work on different semantic levels,which is also not considered sufficiently.In this paper,we design an embedding based method to predict the probability of existing social link between two users with multi-level semantic influence of each neighbor of every user.The main work and innovations of this paper are as follow:(1)With observed neighbor relationships and textual attributes of users,we train a single joint embedding vector for each user with semantic influence of his/her neighbors.Instead of using a constant influence score of a neighbor,we model the special influence of each neighbor toward this user.The influence is modeled according to the textual attributes of the neighbor and this user.Finally,for any pair of users who are not connected in current network,we predict the missing link between pair of users by computing the similarity between the embedding vectors of them.(2)We jointly model the local-level and global-level semantic of neighbor influence in network embedding training.On the one hand,local-level semantic influence portrays the interaction of two users at some particular terms.On the other hand,global-level semantic influence refers to the influence of overall interest of neighbors.(3)In experiments,we reveal how social influence works in the generation of social links.The experimental results on four real-world networks datasets show how the proposed method works better than baseline methods.