<p>Accurate traffic forecasting plays a pivotal role in enabling intelligent transport systems to optimize urban mobility and reduce congestion. However, current traffic forecasting methods primarily model spatial correlations at the node level, such as those between directly connected or neighboring sensors. Comparatively fewer studies explicitly construct region-level correlation graphs from multi-view node relations and jointly integrate them with node-level graphs in a multi-step prediction framework. To overcome these limitations, we propose a recursive multi-granularity graph convolutional network (RMGCN) for traffic prediction. RMGCN comprises three key modules: (1) a graph construction (GC) module that generates multi-granularity graphs at the node and region level; (2) a multi-granularity graph convolutional network (M-GCN) module that adaptively integrates heterogeneous graphs and multi-granularity graph convolution features through trainable kernels, capturing comprehensive spatial correlations. (3) A temporal attention (T-Attn) module that captures the temporal correlations. To capture temporal dependencies across different prediction steps, the proposed recursive framework jointly exploits the multi-granularity spatial representations and hidden-state information at each prediction step, and combines them with the original input to construct the input for the next step, thereby forming a unified end-to-end spatio-temporal learning process. Experimental results demonstrate that RMGCN achieves stable and competitive prediction performance across multiple forecasting horizons, with particularly clear advantages in reducing large prediction errors as reflected by RMSE.</p>