Dynamic Transit Flow Graph Prediction in Spatial-Temporal Network

Traffic flow prediction is of great importance for traffic management. However, most existing researches only focus on region flow or road segment flow (vertex value) prediction, and the transit flow (edge weight) prediction is largely untouched. Compared to region flow and road segment flow prediction, transit flow prediction is more challenging in that 1) the transit flow between pairs of regions has complex spatial-temporal dependencies, and 2) it has larger changes over time due to the large number of region pairs. To address these issues, in this paper we define the transit flow as edges in directed graphs and formulate the transit flow prediction problem as a dynamic weighted link prediction problem. We propose a deep learning based method called Spatial-Temporal Network (STN) to make an accurate prediction of the transit flow. The STN model combines graph convolutional network (GCN) and long short-term memory (LSTM) to capture the dynamic spatial-temporal correlations. To capture the static topological structure, the neighborhood relation graph is adopted as an auxiliary graph to improve the prediction accuracy, and a two-stage-skip strategy is adopted to allow edge features reused which makes the STN focus more on the edge values compared to simple GCN modeling. We conduct the proposed STN model and verify its effectiveness in transit flow prediction on two real-world taxi datasets. Experiments demonstrate that our model reduces the prediction RMSE error by approximately 15.88%–52.48% on real-world datasets compared to state-of-the-art methods.