Modeling of Spatial-Temporal Dependency in Traffic Flow Data for Traffic Forecasting

Accurate traffic forecasting enables efficient traffic management. However, traffic prediction is a challenging task, due to the fact that the traffic system itself is composed of complex interactions of multiple agents (e.g. randomly mixed vehicles with different mechanical characteristics, drivers with diverse driving habits, and road network with unique spatial-temporal correlation attributes among various nodes). To model these interactions, existing methods usually use a given static spatial adjacency graph to represent spatial-temporal dependencies. As a result, they can hardly capture the dynamic spatial-temporal relationships in the traffic network. This paper proposes a new Multi-Head self-Attention based Spatial-Temporal Information Graph Convolutional Networks (MH-ASTIGCN) for traffic flow forecasting where dynamic dependencies between traffic nodes are modelled for predicting the traffic evolution. Firstly, a data-driven strategy for generating temporal information graphs is proposed to amend the spatial correlation that cannot be fully captured by static spatial adjacency graphs. Secondly, we design a novel spatial-temporal attention mechanism in our MH-ASTIGCN, considering the similarity of the traffic patterns between nodes in a road network as a priori, which can not only learn local spatial-temporal dependence, but also capture global deep spatial correlation and temporal features. Thirdly, we use different attention heads in the multi-head attention mechanism to capture the complex multi-scale dependence among the neighborhoods of the spatial graph. Finally, we extend the graph convolutional neural network by integrating our improved attention mechanism to predict traffic flow. Experiments on several real-world data sets show that our proposed method outperforms several recent baseline methods, especially in long-term prediction. We will release our code at https://github.com/SYLan2019/MH-ASTIGCN.