Dynamic multi-graph spatio-temporal learning for citywide traffic flow prediction in transportation systems

Abstract

The complexity and dynamic nature of urban traffic systems necessitate efficient resource management for accurate traffic flow forecasting, enabling real-time adaptation and optimized resource allocation. Recent advancements in multi-graph spatio-temporal graph neural networks (STGNN) have demonstrated their capability to capture spatio-temporal correlations at multiple scales, significantly improving prediction accuracy. However, a persistent challenge lies in effectively aggregating neighborhood information for node representation learning, particularly in scenarios with sparse connectivity. To address this limitation, we propose an Attention-based Dynamic Multi-Graph Module (ADMGM) for traffic prediction, integrating Federated Learning (FL) within a Multi-Access Edge Computing (MEC) architecture. Our approach incorporates an Adaptive Enhancement Module (AEM) deployed at the edge, pre-trained to process high-volume, heterogeneous data from IoT devices. The ADMGM model comprises four key components: closeness, daily, weekly, and an external branch, each contributing to a comprehensive spatio-temporal representation of traffic dynamics. The AEM leverages long-term historical data at each node, capturing inter-node dependencies to generate enriched feature representations while enhancing the model ability to generalize across diverse traffic patterns. Furthermore, we introduce a clustered feature correlation graph to uncover latent relationships within long-term time series data, thereby strengthening spatio-temporal modeling. Extensive experiments on the TaxiBJ and BikeNYC datasets demonstrate that our model significantly reduces prediction errors, achieving state-of-the-art performance in traffic forecasting.