<p>Real-world networks are always dynamic. Dynamic network representation learning represents nodes in a network as low-dimensional, dense, real-valued vectors while preserving semantic and inferential information as much as possible, it has achieved significant success in areas such as link prediction. It focuses on capturing changes in both the spatial dimensions of the network (e.g., the addition and deletion of nodes and edges, and internal network characteristics like behavioral patterns) and the temporal dimensions (e.g., historical influences). Static network representation learning cannot promptly reflect spatial and temporal changes. To address this issue, we propose a sequence generation method based on community adaptive temporal walking (DyCATW) to reflect the internal evolution patterns of networks. Furthermore, in consideration of the impact of historical network data on network evolution, we propose a self-attention mechanism combining community attention and temporal attention modules. This approach generates node walking sequences through DyCATW, which aggregates node neighbor, community, and temporal information for dynamic network representation learning. Experimental results for temporal link prediction on several real-world datasets demonstrate the superior performance of DyCATW, with results by up to 5.1%.</p>