Time-Graph Adjustive Graph Convolutional Recurrent Network for Traffic Forecasting

Abstract

Traffic forecasting is a crucial undertaking in the transportation domain. Current practices rely heavily on Recurrent Neural Networks (RNNs) and Temporal Convolutional Networks (TCNs) to model temporal dependencies in traffic forecasting. However, these approaches tend to overlook the interdependence of multi-hop time steps, impeding their ability to capture long-term dependencies and ultimately limiting their effectiveness in long-term forecasting. To address this issue, we present a novel method, the Time-Graph Adjustive Graph Convolutional Recurrent Network (TAGRN), for traffic forecasting. Our approach employs a Time-Graph model on the temporal domain, treating each time step of the traffic series as a graph node. We incorporate the similarity between time steps as weighted edges, enabling the Time-Graph to capture correlations between multi-hop time steps and model long-term dependencies through graph convolution. We utilize the Simple Graph Convolution (SGC) technique for information propagation due to its simplicity, linearity, and efficiency. Additionally, we introduce an adaptive semantic graph to enhance the capture of spatial information on the spatial domain. Experimental results on various real traffic datasets demonstrate the effectiveness of our proposed method. Compared to existing approaches, TAGRN achieves superior performance in long-term traffic forecasting, highlighting its potential for practical applications. The code is available at https://github.com/sqy123qwer/TAGRN.