In the human skeleton-based action recognition area, most researches aim to extract and fuse the spatial and temporal information by exploiting Graph Convolution Networks (GCNs), which only consider the information propagation of nodes in small neighborhoods, ignoring the impact on the global graph. Therefore, the actions with complex spatial-temporal relationships are not effectively expressed. In this paper, we propose a novel Attention Residual Connection Based Graph Convolution Networks (ARGCNs) to further integrate spatial-temporal features of the global graph with graph attention mechanism and adaptive feature enhancement. Graph attention mechanism focus on identify the key points of every action, since a few joints are more critical than others of the whole body in some action. As the residual connection of ARGCNs, the graph attention mechanism can also increase the flexibility and generalization of model. Experiments demonstrate that the proposed model is effective, and is competitive compared with state-of-the-art methods.