Spatial-temporal graph attention network for traffic flow forecasting

Abstract

Traffic flow forecasting is one of the most challenging tasks in intelligent transportation system, which is very important and critical for many traffic services. Although there are existing works on forecasting the future traffic flow, the majority of them have certain limitations on modelling spatial-temporal correlations: they usually use different modules for spatial and temporal dependencies, and lack abilities of utilizing the historical periodic data. In order to overcome these problems, a novel Spatial-Temporal Graph Attention network (STGAT) is proposed. Considering the periodic characteristic of traffic data, the raw data is transformed into a three-dimensional tensor corresponded to the weekly period and recent period before inputting into the STGAT model. The random walk theory is used to construct a localized spatial-temporal graph to model the spatial-temporal dependencies, and an elaborately designed spatial-temporal graph attention mechanism is used to synchronously capture the complex spatial-temporal dynamic correlations. By calculating the similarity of spatial-temporal features of weekly and recent period input, the model integrates the periodic features into the target. Experimental results on two public real-world datasets demonstrate that STGAT achieves superior performance and consistently outperforms other baselines.