Skeleton-based gesture recognition (SHGR) is a very challenging task due to the complex articulated topology of hands.Previous works often learn hand characteristics from a single observation viewpoint. However, the spatial structure of the hand skeletons cannot be investigated explicitly, and various context information hidden in multiple viewpoints is disregarded. To resolve this issue, we propose a novel multi-view hierarchical aggregation network (MVHANet) for SHGR, which is composed of three parts. Firstly, two-dimensional non-uniform spatial sampling (2DNUSS), a novel strategy forming extrinsic parameter distributions of virtual cameras, is presented to enumerate viewpoints to observe hand skeletons from reasonable directions. Afterwards, we adopt coordinate transformation to generate multi-view hand skeletons and employ a multi-branch convolutional neural networks (CNNs) to further extract the multi-view features of hand skeletons.Furthermore, we exploit a novel hierarchical aggregation network including hierarchical attention architecture (HAA) and global context modeling (GCM) to fuse the multi-view features. Specifically, the HAA is composed of view level and group level for learning the local correlation and discriminative weights of multiple views; GCM is aimed to capture the global information of all the views from each level. Finally, the outputs of HAA and GCM are combined for hand gesture classification. In this way, we fully leverage the multi-view information of hand skeletons and improve the performance of recognition. We conduct extensive experiments on publicly available SHREC'17 Track dataset, DHG-14/28 dataset and First-Person Dynamic Hand Actions dataset. Experimental results demonstrate that our work can be competitive with the state-of-the-art methods.