<p>Skeleton-based human action recognition remains challenging due to the high similarity and overlap between joint movements across different actions, which often leads to misclassification. This paper addresses this problem by proposing a novel approach that integrates action clustering, joint-level voting, and a dual-feature representation to enhance classification accuracy and robustness. The proposed method first divides actions into three meaningful groups—upper body, lower body, and mixed body—based on joint activity, reducing the number of classes considered during classification and simplifying the learning task. Each action is represented using a combination of the joint contribution ratio, which quantifies the importance of each joint, and joint movement direction, which encodes the trajectory of motion. During inference, each joint independently votes for one or more action classes within its cluster, and the final prediction is obtained through a weighted majority mechanism. Evaluated on both the UTD-MHAD and CZU-MHAD datasets, the proposed approach achieves validation accuracies of 96.3% and 95.7%, respectively, outperforming several baseline and state-of-the-art models. The results demonstrate the effectiveness of combining motion-aware joint features with cluster-based voting for improved and interpretable action recognition.</p>