Skeleton-based human action recognition is a critical research topic in computer vision, focusing on extracting and learning discriminative skeletal features to achieve high-precision action classification. However, the presence of ambiguous actions severely impacts classification accuracy. To address these challenges, this paper introduces two key innovations based on data optimization and computational complexity reduction as well as spatio-temporal feature refinement: A Multiple-Input-Branches architecture is employed in the early stage of data processing to facilitate early feature fusion, enabling the model to learn complementary information across different modalities more effectively. This design enhances computational efficiency while improving the model’s generalization ability. To improve the recognition of highly similar actions, an Ambiguous-Feature-Refinement module is proposed to extract distinctive spatio-temporal features. This mechanism enhances the model’s sensitivity to action details, thereby achieving more refined spatio-temporal feature modeling. The proposed Ambiguous Action Recognition under Multi-Input-Branch Skeleton Feature model (GCN+) is evaluated on two large-scale public datasets, ⅅ60 and ⅅ120, covering four single-modal settings as well as their fused modalities. Experimental results demonstrate that: n single-modal settings, GCN+ outperforms the baseline models, achieving a 2.6% accuracy improvement under the X-Sub evaluation protocol on the ⅅ120 dataset, indicating superior robustness in recognizing actions across different subjects in complex environments. In fused-modal settings, GCN+ achieves a 3.2% increase in X-Sub accuracy and a 3.0% increase in X-Set accuracy on the ⅅ120 dataset, further validating its applicability in large-scale data scenarios and its outstanding performance in cross-subject and cross-view action recognition tasks. Overall, the experimental results confirm that GCN+ exhibits strong generalization capability, high computational efficiency, and exceptional performance in ambiguous highly similar actions, providing an effective and robust solution for skeleton-based action recognition in complex environments.