RW-STT: random walk embedded spatial-temporal transformer networks for traffic flow forecasting

Predicting traffic patterns is crucial for urban traffic management and planning, which predicts future traffic data based on current traffic conditions. To acquire better prediction accuracy, traffic flow forecasting requires complex dependencies in both temporal and spatial information. The spatial information is highly related to the traffic network, and how to effectively capture the relationships between graph nodes in the traffic network is crucial. To deal with this problem, we propose to utilize random walk embedding to capture the information transmission and correlations between nodes. Then, we combine random walk with transformer and propose random walk embedded spatial-temporal transformer (RW-STT) model to deal with the traffic flow forecasting problem. Experimental results on three traffic datasets demonstrate that RW-STT can capture more comprehensive correlations between spatial nodes and outperforms other baselines significantly.