In this paper, we study the problem of one-shot skeleton-based actionrecognition, which poses unique challenges in learning transferablerepresentation from base classes to novel classes, particularly forfine-grained actions. Existing meta-learning …