Temporal networks are a natural representation for a wide range of systems, from social interactions to biological processes. Yet, despite significant advances in generative modeling in other domains, the generative modeling of temporal networks remains underdeveloped. This is due in part to their unstructured nature, which imposes prohibitive computational costs for large networks. We present a method that builds on recent advances in machine learning and graph representations to overcome this computational burden. In particular, we show that predicting the evolution of a node’s local topology is sufficient to reproduce realistic macro-level statistics.