Space-Time adaptive network for origin-destination passenger demand prediction

Short-term origin-destination passenger demand prediction involves modeling spatial and temporal characteristics of urban traffic, such as periodicity in demand rate and directionality in flow path. Meanwhile, spatial and temporal heterogeneities often lead to constantly evolving dynamics in in passenger demand, e.g., passengers may exhibit different mobility patterns at different periods or in different regions. Many models fail to capture these heterogeneities and adjust parameters adaptively, leading to suboptimal prediction results. In this paper, we propose a novel space-time adaptive network (STAN) to address these issues. Spatially, an edge-based backbone with a global receptive field is devised. Edge embeddings directly represent pair-wise relations between regions, preserving more fine-grained information and directional interactions. The backbone adaptively updates edge embeddings by fusing static and dynamic information from origin and destination regions, enabling the model to learn intricate spatial relations from simple input data (i.e., basic relation graphs and historical OD matrices). Temporally, a prompter mechanism is proposed to inject temporal information into model parameters, making them timedependent. The parameter values exhibit periodicity and continuity for all periods, meanwhile, they can be adjusted for each specific period. It makes the model time-aware and enables it to identify similar periods and differentiate dissimilar ones during training. Extensive experiments are conducted on two real-world datasets (i.e., ten-month taxi trips in New York and one-month ride-hailing trips in Ningbo), and the results demonstrate that our model outperforms baseline models and automatically learns certain spatial and temporal semantics. With its simple yet highly scalable structure, our model proves beneficial for implementations and can assist related tasks such as driver-passenger matching and surge pricing.