Summary of benchmark datasets.

Accurate traffic flow forecasting is essential for intelligent transportation systems, yet the nonlinear and dynamically evolving spatio-temporal dependencies in urban road networks make reliable prediction challenging. Existing graph-based and attention-based approaches have improved performance but often decouple spatial and temporal learning, which leads to redundant computation and weak directional interpretability. To address these limitations, we propose DSSA-TCN, a unified framework that establishes an alternating spatio-temporal coupling mechanism, where each temporal convolutional block is tightly integrated with an adaptive spatial module that combines sparse attention with diffusion-based graph convolution. Within this mechanism, adaptive sparse attention dynamically selects the most informative neighbors to reduce spatial complexity, and bidirectional diffusion convolution enforces physically consistent directional and multi-hop propagation over the road topology. Temporal patterns are modeled with gated dilated convolutions to preserve parallelism and stability. Comprehensive experiments on six real-world datasets demonstrate that DSSA-TCN achieves superior forecasting accuracy and computational efficiency while providing interpretable spatial reasoning. These results indicate that layer-wise coupling of adaptive sparsity and diffusion within a causal temporal backbone offers a scalable and physically grounded paradigm for spatio-temporal traffic prediction.