- Abstract: We propose GraphGAN - the first implicit generative model for graphs that enables to mimic real-world networks. We pose the problem of graph generation as learning the distribution of biased random walks over a single input graph. Our model is based on a stochastic neural network that generates discrete output samples, and is trained using the Wasserstein GAN objective. GraphGAN enables us to generate sibling graphs, which have similar properties yet are not exact replicas of the original graph. Moreover, GraphGAN learns a semantic mapping from the latent input space to the generated graph's properties. We discover that sampling from certain regions of the latent space leads to varying properties of the output graphs, with smooth transitions between them. Strong generalization properties of GraphGAN are highlighted by its competitive performance in link prediction as well as promising results on node classification, even though not specifically trained for these tasks.
- TL;DR: Using GANs to generate graphs via random walks.
- Keywords: GAN, graphs, random walks, implicit generative models