In action recognition based on the skeleton, graph convolutional networks (GCNs) have shown great superiority based on the design of making the skeleton in the video a spatiotemporal map and extracting features from the spatiotemporal map. However, the topology of the skeleton in GCN-based methods is pre-designed according to prior knowledge, which limits the capacity of the network to learn high-level topology about the skeleton. To improve this deficiency, we design a temporal-difference adaptive graph convolutional network (TDA-GCN) that can learn the potential topology of the human skeleton from the input data, which is augmented using the channel attention module. Experiments show that TDA-GCN achieves state-of-the-art performance on two large-scale skeleton datasets, NTU-RGBD and Kinetics-Skeleton.