DMGI-Net: Dynamic Multi-Graph Integration Network for Spatio-Temporal Traffic Forecasting

Traffic flow forecasting is a foundational component of data-driven traffic management. Although Graph Neural Networks (GNNs) have achieved promising results, most existing methods still rely on static graph structures and struggle to capture spatio-temporal heterogeneity. To overcome these limitations, we introduce DMGI-Net (Dynamic Multi-Graph Integration Network) with three key innovations: (1) a dynamic multi-graph construction strategy that integrates topological, semantic, and flow-based graphs to capture evolving correlations in traffic networks; (2) a convolution-augmented spatio-temporal multihead attention module designed to learn both local variations and long-range global interactions; and (3) a heterogeneous data augmentation (HDA) mechanism that strengthens feature representations through traffic-level memory modeling and graphlevel topology refinement. Comprehensive evaluations on three public traffic benchmark datasets show that DMGI-Net consistently outperforms competitive baseline methods across multiple forecasting horizons, indicating its robustness and effectiveness for practical traffic prediction applications.