Dynamic graphs serve as abstractions of real-world dynamic networks. They represent a concrete and profound restoration of many scenarios in the real-world. For instance, various types of terminal intelligent agents in social networks, recommendation systems, and biological networks facilitate collaborative work within specific group topologies. Despite recent advancements in research on representation learning for dynamic graphs, the factorized representation of features across different dimensions and potential causality have not been adequately considered or explicitly modeled to capture dynamic patterns. The existing literature predominantly relies on manual extraction of temporal and spatial features, which fails to adequately capture the underlying causal relationships. In this study, we propose a novel Dynamic Graph with Spatio-Temporal Disentanglement (DGSTD) that effectively disentangles the spatio-temporal features of the dynamic network within our model. The proposed method sample and sparsely encode the node attribute features under time constraints to find out meaningful structures and patterns for representing graph features, effectively capturing potential spatio-temporal factorized representation. We further used a combination of loss functions to optimize the model. Our approach exhibits distinct advantages in both transductive and inductive settings across four authentic datasets.