Research On Network Aglinment Based On Attention Mechanism And Graph Neural Network

As more and more users are active in multiple social networks at the same time,a large number of researchers have started to pay attention to the research on network alignment,which is beneficial to various machine learning tasks such as cross-domain link prediction,cross-domain recommendation,and cross-domain information diffusion,by using more comprehensive information from multiple social networks.Network alignment refers to connecting two social networks that are not directly related through the user accounts of the same user entity in different social networks.The implicit link relationship built by the same user entity between accounts in two different social networks is usually called anchor link.Most of the existing research on network alignment is based on the user profile information to building an anchor link classification model or directly align network based on the similarity of users' neighborhood structures in different social networks.These methods ignore the unreliability of user profile information in social networks and the heterogeneity of social structure in different social networks,thereby affecting the performance of network alignment.Inspired by the successful application of network embedding representation in a single social network,more and more researchers have begun to apply the method of network embedding to the alignment problem of multiple social networks in order to fully mine the potential information of users in social networks.However,existing network alignment methods based on network embedding representation often compress and simplify the association relationships between users in social networks when modeling complex interaction behaviors of users in social networks,and usually ignore the different influence between users and their neighbors,as well as the influence difference between corresponding features.In recent years,the researches of graph neural network(GNN)has proved its advantages in the learning tasks of mining irregular graph structure information,while the social network happens to be a graph structure composed of user entity associations.In view of this,this thesis studies network alignment based on graph neural network,and by combining attention mechanism,a network alignment method based on double-layer graph attention neural network is proposed.Firstly,we integrate graph neural network with the attention mechanism,and propose a double-layer graph attention neural network model which includes the user graph attention layer and the feature graph attention layer.Among them,the user graph attention layer is used to model different influence between users and their neighbor users,so as to learn the user-level node representation vector;and the feature graph attention layer is used to model the influence differences between different corresponding features of the user and its neighbor users for learning the feature-level node embeddings.Then,we fuse the user node representation vectors which is obtained from different perspectives by gating mechanism and learn the node representation vector for each user in source social network and the target social network.Finally,we propose a bidirectional alignment strategy to align the same user entities between different social networks based on the learned node representation vectors of the two social networks,in order to ensure that the user entities between different social networks meet one-to-one alignment constraints.We use two different types of aligned network datasets to verify the network alignment method proposed in this thesis.The experimental results show that,compared with the existing mainstream network alignment algorithms,the proposed network alignment method based on double-layer graph attention neural network has higher accuracy and better robustness.This thesis firstly introduces the current research status and related technologies of network alignment,and analyzes the problems of existing network alignment methods.Then we elaborate on the network alignment method based on the double-layer graph attention neural network,and give the design and implementation of our scheme.Finally,we analyze the experimental results of the network alignment method on different types of aligned network datasets,and summarize the work of this thesis and the direction of future work.