Traffic Flow Driven Spatio-Temporal Graph Convolutional Network for Ride-Hailing Demand Forecasting

Accurately predicting the demand for ride-hailing in the region is important for transportation and the economy. Prior works are devoted to mining the spatio-temporal correlations between regions limited to historical demand data, weather data, and event data, ignoring rich traffic flow information related to citizens' travel. However, due to the dynamic characteristics of traffic flow and the irregularity of the road network structure, it is difficult to utilize traffic flow information directly. In this paper, we propose a framework called traffic flow driven spatio-temporal graph convolutional network (TST-GCN) to forecast ride-hailing demand. Specifically, we construct a novel region graph based on point of interest (POI) information to model the association between different regions. Besides, we design a stacked traffic-region demand graph convolutional network (TRGCN) module, which is composed of two kinds of nested graph convolutional network structures, effectively modeling the spatial dynamic dependency between regions. Then, the convolution long short-term memory (ConvLSTM) layer is further adopted to obtain spatio-temporal features. We evaluate the proposed model on two real datasets, and the experimental results show that our model outperforms many state-of-art methods.