Abstract: In skeleton-based action recognition, it has been a dominant paradigm to extract motion features with temporal convolution and model spatial correlations with graph convolution. However, ...