SeqST-GAN

Abstract

Citywide crowd flow data are ubiquitous nowadays, and forecasting the flow of crowds is of great importance to many real applications such as traffic management and mobility-on-demand (MOD) services. The challenges of accurately predicting urban crowd flows come from both the nonlinear spatial-temporal correlations of the crowd flow data and the complex impact of the external context factors, such as weather, holidays, and POIs. It is even more challenging for most existing one-step prediction models to make an accurate prediction across multiple future time slots. In this article, we propose a sequence-to-sequence (Seq2Seq) Generative Adversarial Nets model named SeqST-GAN to perform multi-step Spatial-Temporal crowd flow prediction of a city. Motivated by the success of GAN in video prediction, we for the first time propose an adversarial learning framework by regarding the citywide crowd flow data in successive time slots as “image frames.” Specifically, we first use a Seq2Seq model to generate a sequence of future “frame” predictions based on previous ones. Then, by integrating the generation error with the adversary loss, SeqST-GAN can avoid the blurry prediction issue and make more accurate predictions. To incorporate the external contexts, an external-context gate module called EC-Gate is also proposed to learn region-level representations of the context features. Experiments on two large crowd flow datasets in New York demonstrate that SeqST-GAN improves the prediction performance by a large margin compared with the existing state-of-the-art.