Flow Prediction via Multi-view Spatial-Temporal Graph Neural Network

Abstract

AbstractIn recent years, the problem of traffic flow prediction in the urban environment has been widely concerned. However, the traffic flow prediction has not been effectively solved for the next period between the origin-destination region pair. In addition, multiple spatial-temporal traffic dependencies exist between the origin-destination area pairs. In this paper, three types of traffic dependencies between origin-destination region pairs were considered: the same origin dependency, same destination dependency, and transfer to dependency. This paper proposed a spatial-temporal forecasting framework for traffic flow prediction between pairs of urban regions with multi-view graphs. This work mainly considered the construction of spatial-temporal deep learning networks under three kinds of multi-view graphs. Finally, the prediction results under the three dependence relationships are fused to get the final prediction results. Comprehensive experiments on two datasets showed that the proposed framework has very high prediction performance, and outperforms the baseline model by more than 6%.KeywordsMulti-viewSpatial-temporal deep learningFlow prediction