Mest-GAN: Cross-City Urban Traffic Estimation with Meta Spatial-Temporal Generative Adversarial Networks

The conditional urban traffic estimation problem aims to accurately estimate the future traffic status based on the changing local travel demands, which has long been an important issue in urban planning. However, most existing methods require the target city to provide a large amount of traffic data. Once traffic estimation is performed in a "new" city where many urban services and transportation infrastructures are not built and thus no prior data is available, those works would fail due to the lack of data. In this paper, we aim to solve the conditional urban traffic estimation problem in case of data scarcity (i.e., the target city cannot provide any prior data) and tackle the main challenges including (1) knowledge learning from the source and (2) knowledge adaptation without prior traffic data. We propose a novel generative adversarial network Meta Spatial-Temporal Generative Adversarial Network (MestGAN), which can successfully estimate traffic in the target city based on local travel demands without the access to any prior traffic data. To address the first challenge, we learn the latent distribution of travel demands with the inference network, the latent distribution also indicates the diverse spatial-temporal traffic patterns. To solve the second challenge, we use the travel demand data in the target city for adaptation, where the inference network infers a latent code guiding the generator to produce accurate traffic estimations. Extensive experiments on real-world multiple-city datasets demonstrate that our Mest-GAN produces high-quality traffic estimations and outperforms state-of-the-art baseline methods.