Social network alignment: a bi-layer graph attention neural networks based method

The task of social network alignment is to identify the user nodes which are active in multiple social networks simultaneously, thus the information from multiple social networks can be integrated to conduct some downstream tasks. Existing network alignment methods mostly establish anchor link prediction models using user profile information or the structural similarities in social networks, which may not effectively model user nodes, thus affecting the effectiveness of social network alignment. This paper focuses on social network alignment by modeling user nodes based on the structural information of irregular graphs and multi-dimensional user features. Specially, Graph Neural Networks (GNNs) models and bi-layer graph attention mechanism are designed to learn the embedding vectors of user nodes in social networks. First, multi-dimensional user features are comprehensively modeled. Then, a user-layer attention mechanism and a feature-layer attention mechanism based on GNN are respectively designed to learn the embedding vectors of user nodes, and the embedding vectors of user nodes in social networks are learned by designing a gated neural network to automatically learn the weight parameters of the user-layer embedding vectors and the feature-layer embedding vectors. Finally, based on the embedding vectors, a bi-directional alignment strategy is proposed to predict the anchor links between the source and target social networks, thus to ensure which meet the constraints of one-to-one alignment relationship. Simulation experiments based on multiple real social network datasets prove that our proposed method achieves better results in metrics of Accuracy and F1 than the existing mainstream social network alignment methods.