A joint topology-data fusion graph network for robust traffic speed prediction with data anomalism

Accurate traffic prediction is critical for Intelligent Transportation Systems (ITS). However, current deep learning models struggle with two potential issues: 1. difficulty in comprehensively capturing multi-scale spatiotemporal dependencies caused by non-linear and dynamic traffic states; 2. the inherent data non-stationarity and anomalies of traffic information may negatively impact on prediction accuracy. To address these limitations, we propose the Graph Fusion Enhanced Network (GFEN), which is structured around two novel components. First, we introduce a novel Topological Spatiotemporal Graph Fusion (TSTGF) technique that synthetically combines the spatiotemporal graphs derived from traffic data cross-correlation with the essential network topological information. This trainable TSTGF technique explicitly models the combined effect of data distribution and network topology, enabling the extraction and fusion of multi-scale spatiotemporal features more effectively than existing adaptive graph learning models (e.g., AGCRN, GWNET). Second, to mitigate the adverse effects of non-stationary and anomalous traffic data, we propose the Enhanced Data Correction (EDC) technique that employs a hybrid mathematical-transformer architecture to adaptively identify and correct anomalies in historical observations while preserving the underlying spatial feature integrity. Extensive experiments on real-world traffic datasets demonstrate the superiority over state-of-the-art methods in prediction accuracy and exhibit faster convergence rates than recent prediction models.