Abstract: In the study of human motion recognition, millimeter-wave radar technology is regarded as a more effective alternative considering the limitations in video and image performance and the protection of privacy, as it can protect privacy and improve the accuracy of human motion feature recognition. For the sparse point clouds generated by millimeter-wave radar, a novel graph neural network named dynamic graph neural network with MLP and bi-directional gated recurrent units (Dynamic GNN-MB)is designed. A dynamic edge selection function is incorporated into the graph neural network, enabling it to autonomously learn the edge weights between point clouds and extract features. Furthermore, a complete human activity recognition framework is constructed by combining Dynamic GNN with stacked bidirectional gated recurrent units. The effectiveness of the network is validated using a public dataset in experiments. The results show that the Dynamic GNN-MB network model achieves an accuracy of 97.05% in human action recognition, demonstrating a higher recognition rate compared to other network architectures.