Abstract—Human Action Recognition (HAR) has seen significant advancements with Graph Convolutional Networks (GCNs), which effectively model skeletal motion dynamics. In this work, we propose a novel HAR framework that integrates GCNs with an Adaptive Adjacency Matrix for spatial modeling and the Mamba State Space Model (SSM) for temporal feature extraction. This hybrid approach aims to balance accuracy and efficiency. It does so by leveraging the structural expressiveness of GCNs for spatial modeling, while harnessing the sequential modeling power of Mamba-SSM for temporal dynamics. The central goal of this paper is to evaluate GCN-Mamba against our own previous model, ImpSGN (Improved Semantic-Guided Network), with a specific focus on reducing model complexity while preserving recognition accuracy. GCN-Mamba achieves competitive performance with an approximately 72% reduction in parameter count (from 4.0 M to 1.1 M), making it a lightweight yet effective architecture. Our findings highlight the trade-offs between accuracy and efficiency in HAR models and demonstrate the potential of state-space modeling in skeletal action recognition. The code is publicly available on GitHub.
 
Keywords—deep learning, Human Action Recognition (HAR), Graph Convolutional Networks (GCNs), Mamba State Space Models (Mamba-SSM)


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).