Human action recognition is a challenging task due to the complexity of human movements and to the variety among the same actions performed by distinct subjects. Recent technologies provide the skeletal representation of human body extracted in real time from depth maps, which is a high dis-criminant information for efficient action recognition. In this context, we present a new framework for human action recognition from skeleton sequences. We propose extracting sets of spatial and temporal local features from subgroups of joints, which are aggregated by a robust method based on the VLAD algorithm and a pool of clusters. Several feature vectors are then combined by a metric learning method inspired by the LMNN algorithm with the objective to improve the classification accuracy using the nonparametric k-NN classifier. We evaluated our method on three public datasets, including the MSR-Action3D, the UTKinect-Action3D, and the Florence 3D Actions dataset. As a result, the proposed framework performance overcomes the methods in the state of the art on all the experiments.