In this work, we pointed out two key challenges encountered by existing DSNE methods in capturing the evolving structure of network polarization in DSNs. These challenges include the inability to model community boundaries effectively and the lack of inductive ability. To address these challenges, we proposed a novel DSNE method, named POLARDSN. This method learns network polarization through the propagation of node-level CBs and enhances inductive ability by incorporating structural isomorphism and temporal transitivity between nodes. We conducted extensive experiments on four real-world dynamic signed network datasets, comparing the performance of POLARDSN against 12 existing NE methods across three experimental settings. The results demonstrate that (1) POLARDSN consistently outperforms all these competitors in terms of dynamic edge prediction accuracy and (2) our design choices are all effective for dynamic edge prediction tasks.Â