Despite the notable success of Graph Convolutional Networks (GCNs) in skeleton-based action recognition, their performance often depends on large volumes of meticulously labeled data, which are frequently scarce. To address this critical limitation, we propose a novel "data and label" efficient GCN model based on active sample learning. Our work makes two primary contributions: First, we introduce a principled probabilistic framework that proactively designs informative unlabeled exemplars for labeling, moving beyond passive selection by jointly considering representativeness, diversity, and model uncertainty. Second, we develop reversible and stable GCN architectures that facilitate an effective mapping between the input and latent data spaces, enabling a more robust and efficient generation of highly relevant exemplars for training. Extensive evaluations on challenging skeleton-based action recognition benchmarks reveal significant improvements achieved by our label-efficient GCNs compared to prior work, demonstrating their superior performance with reduced annotation burden.