Predicting Citywide Crowd Flows in Critical Areas Based on Dynamic Spatio-Temporal Network

Citywide crowd flow prediction is an important problem for traffic control, risk assessment, and public safety, especially in critical areas. However, the large scale of the city and the interactions between multiple regions make this problem more challenging. Furthermore, it is impacted by temporal closeness, period, and trend features. Besides, geographic information and meta-features, such as periods of a day and days of a week also affect spatio-temporal correlation. Simultaneously, the influence between different regions will change over time, which is called dynamic correlation. We concentrate on how to concurrently model the important features and dynamic spatial correlation to increase prediction accuracy and simplify the problem. To forecast the crowd flow in critical areas, we propose a two-step framework. First, the grid density peak clustering algorithm is used to set the temporal attenuation factor, which selects the critical areas. Then, the effects of geographic information on spatio-temporal correlation are modeled by graph embedding and the effects of different temporal features are represented by graph convolutional neural networks. In addition, we use the multi-attention mechanism to capture the dynamic spatio-temporal correlation. On two real datasets, experimental results show that our model can balance time complexity and prediction accuracy well. It is 20% better in accuracy than other baselines, and the prediction speed is better than most models.