Abstract: Dynamic graphs widely exist in practical application scenarios, and their representation learning can be effectively applied to various downstream analysis tasks. The topological structures and node features of dynamic graphs are evolving, and there are structural correlations and feature correlations between nodes. Existing research ignores the impact of the above two types of correlation on dynamic graph representation learning. To address this problem, a dynamic graph representation learning model based on a joint attention mechanism(DyGRL-JAM) is proposed in this paper. DyGRL-JAM learns structural correlations from dynamic topological structures and combines them with feature correlations to construct the joint attention. Then, a graph tensor convolution is adopted to implement node spatiotemporal message passing to obtain node representation. Experiments are conducted on six dynamic graph datasets, and the results show that DyGRL-JAM has better representation learning capabilities for dynamic graphs than state-of-the-art methods.