Wide Area Network (WAN) research benefits from the availability of realistic network topologies, e.g., as input to simulations, emulators, or testbeds. With the rise of Machine Learning (ML) and particularly Deep Learning (DL) methods, this demand for topologies, which can be used as training data, is greater than ever. However, public datasets are limited, thus, it is promising to generate synthetic graphs with realistic properties based on real topologies for the augmentation of existing data sets. As the generation of synthetic graphs has been in the focus of researchers of various application fields since several decades, we have a variety of traditional model-dependent and model-independent graph generators at hand, as well as DL-based approaches, such as Generative Adversarial Networks (GANs). In this work, we adapt and evaluate these existing generators for the WAN use case, i.e., for generating synthetic WANs with realistic geographical distances between nodes. We investigate two approaches to improve edge weight assignments: a hierarchical graph synthesis approach, which divides the synthesis into local clusters, as well as sophisticated attributed sampling. Finally, we compare the similarity of synthetic and real WAN topologies and discuss the suitability of the generators for data augmentation in the WAN use case. For this, we utilize theoretical graph metrics, as well as practical, communication network-centric performance metrics, obtained via OMNeT++ simulation.
