The skeleton-based action recognition technology can effectively avoid the background interference and occlusion problems in the image. However, the recognition of similar actions is still a challenge. In this paper, a multi-scale dynamic topological modeling method (MDTM) is proposed to solve this problem. The topological modeling through the convolution kernel generated from the original data, increases the connection of the convolution process to the original data compared to the previous randomly generated ones, effectively distinguishing similar actions. In addition, MDTM uses a multi-scale temporal convolutional network to obtain a wider receptive field, which can effectively extract the temporal information in the action. At the same time, a dynamic topology learning method is utilized to design a spatiotemporal information extractor that can effectively extract the spatiotemporal information in the action to dynamically adjust the topology structure. Extensive experiments have performed on three large-scale datasets, NTU RGB + D 60, NTU RGB + D 120, and NW-UCLA to validate the effectiveness of MDTM. The results show that MART-GCN performs better than the others in terms of accuracy and number of parameters.