<p>Link prediction in dynamic graphs models real-world dynamic networks, providing a concrete and insightful representation of various scenarios. Despite recent advancements in dynamic graph learning, the factorized representations of features across different dimensions and potential causality have not been sufficiently considered or explicitly modeled to capture dynamic patterns for link prediction. In this study, inspired by the variational autoencoder, we propose a <i>variational autoencoder-based spatio-temporal disentanglement</i> for link prediction in dynamic graph that effectively disentangles the spatio-temporal features of the dynamic network within our model. We separate spatial features and continuous time features from the graph data, anonymously encode and sparsely represent the temporal features to enhance the method’s accuracy. We then implement a spatio-temporal features disentanglement method that effectively captures potential spatio-temporal factorized representations to identify meaningful structures and patterns for spatio-temporal correlation. Additionally, we utilize an integrated loss function to optimize both temporal and spatial losses, enabling our model to adapt to complex spatio-temporal dynamics. Our approach demonstrates state-of-the-art in spatio-temporal link prediction across four authentic datasets in both transductive and inductive settings.</p>