Augmenting Spatio-Temporal Dependencies for GNN-Based Short-Term traffic Flow Prediction

Accurate traffic prediction has drawn increasing attention in recent years due to its potential to enhance traffic management measures. Most state-of-art studies employ graph representative learning and time series models to predict traffic by jointly capturing the spatial and temporal dependencies of traffic patterns. However, the misinterpretation of pattern similarity between road segments caused by the spatio-temporal lag effect is widely overlooked in existing studies, potentially leading to incorrect spatial propagation of traffic information in Graph Neural Networks (GNN). To capture this pattern, a hybrid framework, named Dynamic Time Warping based Spatio-Temporal Graph Convolution Network (DTW-ST-GCN), is proposed. Firstly, this framework utilizes Dynamic Time Warping (DTW) to align the time series data and capture the hidden spatio-temporal correlations, which are then used to generate sequence-based matrices. Secondly, the traffic volume data of the road segments are separately processed by Graph Convolution Network (GCN) and Recurrent Neural Network (RNN), serving as spatial and temporal encoders, respectively. Then, the outputs from both encoders are fused and fed to an RNN decoder, which enhances the model’s effectiveness in capturing the temporal dependencies. The proposed model is evaluated through carefully designed experiments using real-world data. Experimental results demonstrate improved model performances and indicate a certain level of robustness against the impact of the significant long-term traffic pattern disruptions.