Title: Efficient Synthetic Network Generation via Latent Embedding Reconstruction

Network datasets are prevalent across disciplines ranging from the social sciences and biology to information systems. The ability to generate realistic synthetic network data holds significant promise for applications spanning network simulation to scientific discovery. However, current black-box generation methods often struggle with practical limitations: they tend to overfit observed data, neglect essential structural characteristics, and demand high computational resources as scale increases. To address these issues, there is a pressing need for synthetic generation techniques that are both computationally efficient and capable of accurately capturing network structures.

This paper presents SyNGLER (Synthetic Network Generation via Latent Embedding Reconstruction), a general and efficient framework grounded in latent space network models. The SyNGLER process begins by analyzing an observed network to learn low-dimensional latent node embeddings through a latent space model. Subsequently, it reconstructs the latent space by employing a distribution-free generator based on these embeddings. To generate synthetic data, SyNGLER samples or resamples node embeddings from this generator within the latent space, which are then used by the latent space network model to produce the final synthetic networks.

By leveraging this latent space approach, SyNGLER maintains distinct network features, including sparsity and heterogeneity in node degrees, while offering more efficient training with reduced computational costs compared to many contemporary deep learning architectures. We establish theoretical validity by deriving consistency results regarding the distance between the true edge distributions and those of the synthetic networks. Empirical evaluations confirm SyNGLER’s effectiveness, showing that it efficiently generates networks that better retain key structural attributes—such as degree distributions and network moments—than existing methods. The source code is publicly available at https://github.com/FeifanJiang/syngler.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC