<p>Dynamic facial expression recognition in the wild is challenged by non-frontal poses, local occlusions, head movements, and a large proportion of expression-irrelevant frames, under which subtle but informative expression patterns are easily submerged. To address these issues, we propose a Spatio-Temporal Adaptive and Region-Aware Transformer (STAR-Former). On the spatial side, STAR-Former adopts a dual-path architecture: a global path employs a Transformer to model holistic facial semantic dependencies, while a local path introduces a hierarchical parent–child attention mechanism to learn soft region masks, emphasizing discriminative local muscle areas. A gating-based fusion module is further used to adaptively integrate global and local features, enhancing the robustness of spatial representations under pose variations and occlusions. On the temporal side, STAR-Former derives frame-wise importance weights from the saliency of parent regions and injects them as attention biases into a temporal Transformer, guiding the model to focus on highly discriminative key frames and suppress responses to expression-irrelevant frames, thereby achieving region-driven end-to-end temporal modeling. Experiments and visualizations on three public benchmarks, DFEW, FERV39k, and AFEW, demonstrate that STAR-Former attains superior recognition accuracy and robustness compared with existing mainstream methods.</p>