Research On User Alignment Methods Across Social Network Fusing Global Information

With the rapid development of social networks,more and more users choose to participate in different social platforms and register social accounts at the same time.How to identify the social accounts belonging to the same user among these different social platforms is called network alignment or user identification,which has become a research hotspot in complex network analysis.Cross-social network user identification plays an important role in many existing learning tasks,such as cross-network recommendation,link prediction,cross-network information transmission,information fusion,etc.With the development of network representation learning in network alignment,some current network alignment methods map nodes in the network to low-dimensional vector space through network embedding and then learn a supervised or unsupervised user alignment model to realize user identification across the social network.However,most of the alignment methods mainly make use of local structural information,and:(1)these researches on user alignment across social networks based on network embedding treat the importance of nodes in the network equally when extracting account features,without distinguishing the different node states of different nodes;Meanwhile,(2)these methods all extract local structural proximity from the context of nodes,which focusing on the micro-structure of the network while ignoring some typical attributes of social networks,such as community structure.To solve the problems above,this thesis mainly proposes the following research methods to solve the problem of user alignment across social networks based on the network topology information in social networks.The main research contents include:1.To solve the problem(1),a cross-social network user alignment method combining node state information is proposed.In addition,node status can well reflect a user's social status on the social platform and reflect the importance of the user in the network.In order to make use of the node state information,this thesis uses the network embedding method to learn the vector representation via fusing the node neighbor relationship and node state information and obtains the low-dimensional vector space that retains the original network structure information.Then,given a set of anchor nodes with known markers,a backpropagation neural network is applied to learn a stable cross-social network mapping function in two vector spaces to identify users aligned between networks.In order to verify the effectiveness of the proposed method,some validation experiments are performed on two real data sets.The experimental results show that the proposed method can effectively improve the accuracy in predicting aligned users.2.For problem(2),a cross-social network user identification method with community structure is proposed.In feature extraction of nodes,most of the existing network structure-based methods focus on extracting the local structure proximity from the local context of nodes,but largely ignore the inherent community structure of social networks.Community structure is one of the most significant features of social networks.Users are usually closely connected with users in the same community,but the connections between users in different communities are relatively sparse.However,neighbors and friends who are usually closely connected naturally form a community.Users who have similar neighbors in the network and exist in similar community structures have a high probability of alignment.Therefore,this thesis introduces community structure into the research of user alignment across social networks and proposes a new model of user alignment based on community retention.Experimental results on real data sets show that this method is superior to other user alignment methods that do not consider community information.To sum up,this thesis mainly studies how to extract account association information from different perspectives by using the local and global characteristics of accounts in social networks,obtain low-dimensional and effective vector representation by using rich user information based on network embedding,and then achieve user identity alignment across social networks according to account characteristics.At the same time,this thesis has carried out experiments on real data sets of different scales,and the validity of the research content in this thesis can be concluded from the experimental results.