A novel generative corrective network structure for traffic forecasting.

Traffic forecasting plays a critical role in intelligent transportation systems aiming to accurately estimate future short-term or long-term traffic conditions. The utilization of neural network-based methods for traffic forecasting has demonstrated significant success. Those models aggregate spatial and temporal information from historical traffic records. The final output module of these models typically utilizes either a fully connected network to directly produce the results or a recurrent neural network to generate results step by step. However, these approaches fail to consider the spatial-temporal dependencies between predicted future traffic conditions, which hinders their forecasting performance. To address this limitation as well as improve forecasting accuracy, we propose a novel traffic prediction framework called the Generative Corrective Network. This framework consists of two procedures: a generative model that captures the spatial temporal relationships inherent in historical traffic condition and generates initial predictions, and a corrective model that amends and updates previous results by considering the relationships among future traffic data to be predicted. To implement this generative corrective network, we design an instantiation utilizing adaptive and dynamic graph convolutional networks with gated recurrent units networks (ADGCN-GRU) based on encoder–decoder structure for the generator, and a local asynchronous attention model conducted on the the representation learning of space and time information and the preliminary results for the corrector. Experiments conducted on two real-world traffic datasets demonstrate that our model can effectively improve forecasting performance compared to existing neural network-based methods. [ABSTRACT FROM AUTHOR]