Recently skeleton-based action recognition has made significant progresses in the computer vision community. Most state-of-theart algorithms are based on Graph Convolutional Networks (GCN), and target at improving the network structure of the backbone GCN layers. In this paper, we propose a novel mechanism to learn more robust discriminative features in space and time. More specifically, we add a Discriminative Feature Learning (DFL) branch to the last layers of the network to extract discriminative spatial and temporal features to help regularize the learning. We also formally advocate the use of Direction-Invariant Features (DIF) as input to the neural networks. We show that action recognition accuracy can be improved when these robust features are learned and used. We compare our results with those of ST-GCN and related methods on four datasets: NTU-RGBD60, NTU-RGBD120, SYSU 3DHOI and Skeleton-Kinetics.