Dstagnn: Dynamic spatial-temporal aware graph neural network for traffic flow forecasting

Authors	Shiyong Lan, Yitong Ma, Weikang Huang, Wenwu Wang, Hongyu Yang, Pyang Li
Title	Dstagnn: Dynamic spatial-temporal aware graph neural network for traffic flow forecasting
Publication	Proceedings of the 39th International Conference on Machine Learning (ICML 2022)
Volume	162
Issue	x
Pages	11906-11917
Year	2022
DOI	x

Introduction

Background

• With a growing number of vehicles in road networks, there is increasing pressure on traffic management systems.
• The development of ITS (Intelligent Transportation System) is urgently needed for efficient traffic management.

Previous Research

• Traffic flow prediction plays a key role in ITS, as it is a necessary prerequisite for the implementation of an intelligent traffic management system.
• The traffic at each recording point (also called node in the road network) presents patterns of highly dynamic and complex temporal-spatial dependency.
• Deep learning methods have become a popular choice for traffic flow prediction from high-dimensional spatial-temporal data.

Proposed Model

• We propose a novel neural network framework, DSTAGNN (Dynamic Spatial-Temporal Aware Graph Neural Network) for traffic flow forecasting, which can capture both short-range and long-range spatial-temporal correlations of the road network.
• We construct a novel graph to capture dynamic attributes of spatial association among nodes by mining from their historic traffic flow data directly, without using a predefined static adjacency matrix.
• We design a new spatial-temporal attention module to exploit the dynamic spatial correlation within multi-scale neighborhoods based on multi-order Chebyshev polynomials in GCN (Graph Convolutional Network).
• An improved gated convolution module is designed, which can further enhance the awareness of the model to dynamic temporal dependency within the road network, via fusing temporal features of multi-receptive fields with multi-scale gated convolution.
• Extensive experiments on real road traffic data sets demonstrate the improved performance of our proposed algorithm, as compared with several baselines including the state of the art algorithms.

Significance

• Our proposed algorithm can capture both short-range and long-range spatial-temporal correlations of the road network, and demonstrate improved performance as compared with several baselines including the state of the art algorithms.

Proposed Model

• It is composed of stacked ST (Spatial-Temporal) blocks and a prediction layer.
• The output of each ST block is concatenated and then sent to the prediction layer in a manner similar to residual connection.
• The model extracts more accurate spatial dependencies between nodes in the road network using historic traffic data.
• It uses a data-driven strategy to formulate the degree of spatial association among nodes directly from historic traffic data, called STAD (Spatial-Temporal Aware Distance), and STAG (Spatial-Temporal Aware Graph).
• The model transforms a set of multi-day traffic data into a probability distribution using the daily traffic volume information at each recording point.
• The conversion cost of each probability mass is obtained using the cosine distance between traffic flow vectors as a cost function.
• The spatial-temporal aware distance is calculated as a matrix A_STAD ∈ RN×N that represents the degree of relevance between the recording points.
• Spatial-Temporal Relevance Graph (A_STRG) is obtained as prior knowledge to supplement the attention P learned from the spatial-temporal attention module.

Experiment

• DSTAGNN achieved the best results in all indicators on four data sets, indicating its excellence compared to baseline methods.
• The spatial-temporal aware distance in DSTAGNN helped the model capture spatial dependency among nodes, allowing it to be applied even in the absence of spatial prior information.
• The proposed spatial-temporal attention mechanism in DSTAGNN better captured dynamic changes in data, resulting in significantly improved prediction performance.
• DSTAGNN responded more quickly and accurately to dynamic changes in peak traffic and recovered faster and maintained higher accuracy in the case of missing data compared to the baseline method.
• Ablation experiments showed that each component in DSTAGNN was effective, as demonstrated by better performance compared to variants that removed certain components.