Skeletal data plays a pivotal role in human action recognition due to the compact form of pose information inherent in the skeletal structure. Graph Convolutional Networks (GCNs) along with attention mechanisms are showing good results for skeleton-based activity classification making use of the natural graph representation of the skeleton. However, such combinations of GCN and attention often fail to capture the spatial relationships between the body parts. We propose a dual-stream architecture integrating Spatial and Temporal Transformer models for skeleton-based activity classification. The proposed approach leverages GCNs and Attention Mechanisms to model spatial relationships between skeletal joints and temporal dynamics in action sequences. Additionally, we introduce a Feature Refinement Head using contrastive learning to enhance discriminative feature representations. We experiment with challenging benchmark datasets, and compare against the recent state-of-the-art, for skeleton-based human action recognition. The proposed method outperforms the recent methods. The codes are available at this link .