Federated learning-based regularized adversarial graph embedding for cross-social network user alignment

The goal of cross-social networks user alignment is to identify corresponding nodes belonging to the same entity across different social networks, which has emerged as a critical focus in various social networking applications. In practice, social platforms often refrain from sharing user information and network structures due to commercial considerations, posing significant challenges to user alignment. To address this problem, we propose a graph embedding method based on federated learning for user alignment. Specifically, each platform first trains models locally to embed its network into a shared latent space, then uploads these models to a third-party server for aggregation. To compensate for performance loss incurred by privacy protection, we introduce adversarial regularization to match a prior Gaussian distribution, thereby enhancing the model’s generalization capabilities. Experimental results on two real social network demonstrate that our approach achieves superior performance compared to current benchmark methods.