GCNoT: A Lie-Algebra-Inspired View-Invariant Graph Model for 3D Skeleton Action Recognition

Skeleton-based action recognition models often suffer from poor generalisation under viewpoint distribution shifts. In this work, we propose GCNoT, a graph-based model that incorporates an explicit rotational inductive bias for view-invariant representation learning. Our method augments a standard GCN backbone with a learnable transformation module T(.) based on Lie-algebraic rotation operators, together with a lightweight selector that adapts rotation contributions across learned angular partitions. Unlike prior approaches, GCNoT does not rely heavily on view-specific data augmentation to achieve robustness. Experiments on NTU RGB+D 60 and 120 show that GCNoT achieves competitive or superior performance compared to recent state-of-the-art methods, while using approximately half the parameter budget of strong baselines such as InfoGCN. Under viewpoint distribution shifts, our model demonstrates substantially improved generalisation, hence highlighting the importance of explicitly encoding geometric structure in graph-based models for robust skeleton-based action recognition.