Traffic Spatial-Temporal Prediction Based on Neural Architecture Search

Abstract

Traffic spatial-temporal prediction is essential for intelligent transportation systems. However, the current approach relies heavily on expert knowledge and time-consuming manual modeling. Neural architecture search can build models adaptively, but it is rarely used for traffic spatial-temporal prediction, nor is it designed specifically for traffic spatial-temporal feature. In response to the above problems, we propose neural architecture search spatial-temporal prediction (NASST), which is a method to automatically generate a traffic spatial-temporal prediction network by performing a differentiable neural network architecture search in an optimized search space. First, we adopt a differentiable neural architecture search method to continuously relax the discrete traffic spatial-temporal prediction model architecture search, and adopt a fusion strategy of comprehensive concatenate and addition (CA) to achieve efficient neural architecture search. Second, we optimize the search space and introduce a series of classic traffic spatial-temporal feature extraction modules, which are more in line with the architectural requirements of traffic spatial-temporal prediction network. Finally, our model is validated on two public traffic datasets and achieves the best predictions. Compared with traditional manual modeling methods, our method can realize the automatic search of high-precision predictive model architectures, which improves the modeling efficiency.