Hetero-STAN: Crowd Flow Prediction by Heterogeneous Spatio-Temporal Attention Network

Abstract

Crowd flow prediction has involved in extensive applications like intelligent transportation and public safety, especially in metropolis where the crowd flow usually show high nonlinearities and complex patterns. Among the existing prediction methods, most of them suffer from (1) implicit long-term spatial dependency, (2) the external factors lack of crucial spatial attributes, (3) complex spatio-temporal dynamics with uncertain external conditions, which yield limited performance. This paper proposes a novel method using spatio-temporal attention network with heterogeneous feature enhancement. Specifically, heterogeneous feature enhancement introduces spatial mapping and Periodic Dilated Convolution (PDC), the former provides the dimension supplement of external factors while PDC could capture the correlations of both spatial and temporal domain. Moreover, a Spatio-Temporal Attention (STA) mechanism is proposed to further obtain the dynamic spatial-temporal correlations. Our framework is evaluated on several citywide crowd flow datasets, i.e. TaxiBJ, MobileBJ and TaxiNYC, the experimental results indicate the proposed method outperforms the state-of-the-art baselines by a satisfied margin.