Deep spatial-temporal networks for crowd flows prediction by dilated convolutions and region-shifting attention mechanism

摘要

Flow prediction at a citywide level is of great significance to traffic management and public safety. Since deep learning has achieved success to deal with complex nonlinear problems, it has drawn increasing attention on making crowd flows prediction through neural networks. Generally, convolutional neural network (CNN) and recurrent neural network (RNN) have been applied to model the spatial-temporal dependency of the city. However, there are still two major challenges in predicting flows. First, it is difficult to train the model with the ability to capture both the nearby and distant spatial dependency by deep local convolutions. Second, daily and weekly patterns in temporal dependency are not strictly periodic for their dynamic temporal shifting in each region. To address these issues, we propose a novel deep learning model which called Local-Dilated Region-Shifting Network (LDRSN). LDRSN combines local convolutions with dilated convolutions to learn the nearby and distant spatial dependency. Furthermore, a new region-level attention mechanism is proposed to model the temporal shifting which varies by region. In the experiments, we compare the proposed method with other state-of-the-art methods in two real-world crowd flows datasets. The Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE) were used as the evaluation indexes. The experiment results show the effectiveness of the proposed model.