On region-level travel demand forecasting using multi-task adaptive graph attention network

Abstract

Accurate travel demand forecasting at the regional level benefits to urban traffic management and service operations. Irregular regions can be naturally represented by graphs, and thus, graph neural network (GNN) is rapidly becoming the mainstream method in region-level travel demand forecasting. Most existing GNN-based methods encode the association between regions into single or multiple fixed adjacency matrices. The limitation of these methods is that they ignore that the correlation between regions may vary over time. Several studies have attempted to capture such change with purely data-driven dynamic graph approaches; however, they are highly dependent on data quality. Furthermore, regions with different functional distributions exhibit diverse demand patterns, which are not fully considered in demand forecasting. To address these issues, we propose a multi-task adaptive recurrent graph attention network, in which the spatio-temporal learning component combines the prior knowledge-driven graph learning mechanism with a novel recurrent graph attention network to capture the dynamic spatiotemporal dependencies automatically. Demand forecasting for all regions is divided into different learning tasks based on region function distributions, and an uncertainty-based multi-task learning component is then developed to coordinate multiple learning tasks adaptively. Experimental results based on two real-world datasets show the superiority of the proposed method in prediction accuracy. Compared with the state-of-the-art dynamic graph methods, the proposed method further reduces prediction error by 4%. In addition, various ablation experiments demonstrate the effectiveness of different components in the proposed method.