Traffic Flow Forecasting based on Adaptive Graph Structure for Spatio-Temporal Attention Mechanism Network

Abstract

With the diversification of information collection, traffic flow data has been expanded from specific areas to large-scale networks, which are represented as spatiotemporal sequence data. Recent research on spatiotemporal graph neural networks (ST-GNNs) has demonstrated excellent performance in capturing these correlations. However, most ST-GNNs rely on predefined graph adjacency matrices or fail to utilize valuable traffic information in node attributes to derive better graph structures. We propose an adaptive graph structure for spatiotemporal attention mechanism network (ASTAMN) model. The learning module for graph structure is based on node properties and constructs the ideal graph structure through weight fusion. The designed spatiotemporal convolution is applied to capture complex spatiotemporal correlations. Experiments on real data sets confirm the effectiveness and robustness of our model.