STLGCN: Spatial-Temporal Graph Convolutional Network for Long Term Traffic Forecasting

As an essential part of intelligent transportation, accurate traffic forecasting helps city managers make better arrangements and allows users to make reasonable travel plans. Current mainstream traffic forecasting models are developed based on spatial-temporal graph convolutional neural networks, in which appropriate graph structures must be generated in advance. However, most existing graph generation approaches learn graph structures based on local neighborhood relationships of urban nodes, which cannot capture complex dependencies over long spatial ranges. To solve the above problems, we propose Spatial-Temporal Graph Convolutional Neural Network (STLGCN) for long-term traffic forecasting, in which a novel graph generation method is developed by measuring multi-scale correlations among vertices. Meanwhile, a new graph convolution method is proposed for extracting valuable features and filtering out the irrelevant ones, which significantly optimizes the process of spatial information aggregation. Extensive experimental results on two real public traffic datasets, METR-LA and PEMS-BAY, demonstrate the superior performance of our algorithm.