A Spatial–Temporal Multitask Collaborative Learning Model for Multistep Traffic Flow Prediction

Abstract

Traffic flow prediction is a fundamental capability for successful deployment of intelligent transportation systems. Traditionally, multiple related prediction tasks are undertaken individually, without considering the relationships among the tasks. This paper presents a spatial–temporal multitask collaborative learning model for multistep traffic flow prediction. The novel approach learns multiple related prediction tasks collaboratively by extracting and utilizing appropriate shared information across tasks. First, each traffic flow prediction problem is formulated as a supervised machine-learning task. Next, the sparse features shared across multiple tasks are learned by solving a regularized optimization problem. To deal with the non-convex and non-smooth challenges, the optimization problem is then transformed into an equivalent convex problem. Finally, the global optimal solution of the convex problem is found by solving a variation of this problem using an alternating minimization algorithm. The proposed model incorporates both the spatial correlation between different observed stations and the intrinsic relationship between different traffic flow parameters, as well as the coarse-grain temporal correlation between different days in a week and the fine-grain temporal correlation between different prediction steps. Application of the proposed model to a real case study for SR180-E freeway in Fresno, California showed its effectiveness, robustness and advantages for multistep traffic flow prediction.