Research On User Identity Linkage Technology Across Social Networks

As an important part of the research on social networks,user identity linkage across social networks plays an important role in multi-source data fusion,information dissemination across networks,cyberspace security management.In recent years,many researchers have paid more and more attention to the research on user identity linkage and many excellent algorithms based on network topology has emerged in endlessly.However,there are still some problems unsolved in the existing research,such as incomplete user feature extraction,inaccurate relationship modeling between users and so on.We focus on using the network topology to solve the problem of user identity linkage across social networks.The main work and innovations are as follows:1.In term of the problem of incomplete user feature extraction,caused by only using proximity structures to learn user feature representations in existing research,a method of user identity linkage across social networks via community preserving network embedding is proposed.Firstly,this method considers both proximity structure and community structure of the social network simultaneously to capture the structural information conveyed by the original network as much as possible when learning the feature vectors of nodes in social networks.Secondly,this method trains the back-propagation neural network to learn a stable nonlinear cross-network mapping function for identities linkage with the known anchor link as supervision information.A series of experiments conducted on real social network datasets and synthetic datasets show that this algorithm not only improves the accuracy of user identity linkage.For example,0)(84)4)9)@1 exceeds 45%,which is more than 30% higher than the most advanced algorithm Deep Link.And this algorithm also has excellent performance with only a few known anchor links.For example,when the known anchor links is only about 10%,0)(84)4)9)@30 can reach more than 50%.2.Focusing on the problem that the traditional similarity measurement learning cannot fully mine the complex implicit relationships between users across social networks in the existing research,a neural tensor network-based method for user identity linkage is proposed.Firstly,this method applies the Random Walks to sample the network structure and use the Skip-gram model to learn the feature representation of users,which maps the original network space to the lowdimensional vector representation space.Secondly,the neural tensor network is used to model the implicit relationship between users across social networks.Finally,this method transforms the problem of user identity linkage into a binary classification problem through a multi-layer perceptron model.Proved by sufficient experiments,the algorithm is more than 70% in precision,recall and F1,which are higher than the baselines.Especially,the F1-value of the algorithm is improved by more than 20% compared with previous work.3.Aiming at the problem that the existing research cannot capture the global structure information of social networks using the shallow network representation learning and the existing research ignores the differences of mutual influence between users and their neighbors,a user identity linkage method based on graph attention network is proposed.This method applies attention mechanism to model the differences of mutual influence between the user and its neighbors.And then this method performs weighted aggregation on the feature information of the user's neighbors to update its own feature representation.The continuous propagation and aggregation of node feature information in social networks could capture the global structural feature.The experiments conducted on the Twitter-Foursquare dataset proves that the method has improved in Precision,Recall and F1 in different degrees,for example,the mothed is about 10% higher than the best baseline.