Deep Flexible Structured Spatial–Temporal Model for Taxi Capacity Prediction

Abstract The prevalence of taxi-hailing applications has brought great convenience to urban travel. People can not only get a taxi anytime and anywhere, but also make an appointment for taxi in advance. Therefore, people are concerned about the current taxi capacity (i.e. the number of vacant taxis) around them. Meanwhile, they also want to know the future capacity to help them choose the appointment time to avoid congestion and plan their itinerary. However, most of the exiting studies only aim to help taxi companies to schedule traffic resources, and can not consider future travel plans for users. In this paper, we propose the Deep Flexible Structured Spatial-Temporal Model (DFSSTM) to tackle the task. In order to explore more sufficient temporal relationship of data, DFSSTM models the temporal dynamics as three views: period, trend and closeness. Then the Siamese Spatial-Temporal Network (SSTN) is designed for each view, which introduces the siamese architecture to capture the spatial–temporal dependencies of inflows and outflows simultaneously. Finally, DFSSTM automatically weights each view and fuses the outputs of the three views to get the final prediction. Experimental results on real-world datasets show that the proposed approach outperforms state-of-the-art methods.