LAUSR

Graph learning-based algorithms are becoming the prevalent traffic prediction solutions due to their capability of exploiting non-Euclidean spatial-temporal traffic data correlation. However, current predictors primarily employ heuristically constructed static traffic graphs in forecasting, which may not describe the latent traffic dynamics well. Existing attempts on dynamically generated traffic graphs also face challenges like prolonged model training time and undermined model expressibility. In this paper, a novel data-driven graph construction scheme based on graph adjacency learning is proposed for graph learning-based traffic predictors. The proposed scheme explores inter-time-series dependency with the graph attention mechanism to embed the sensor correlation in a latent attention space, which determines the correlation of any possible sensor pairs for traffic graph construction. Comprehensive case studies on three real-world traffic datasets reveal that the proposed scheme outperforms state-of-the-art static and dynamic graph construction baselines. Additionally, time-varying and sparse graph construction schemes are devised and assessed to boost the efficacy, and a hyper-parameter test develops guidelines for parameter and model architecture selection.