<p>Online Social Networks (OSNs) have become the main platform for individuals to share information, but the rapid dissemination of information has also resulted in an increase in the spread of rumors. Some structure-based deep learning methods focus on the depth of propagation while disregarding the wide disperse patterns of rumors. They are also based on static graphs, which are incompatible for capturing the dynamic nature of real-world rumor propagation. To overcome these challenges, we propose a novel Bi-directional Temporal Graph Attention Networks, termed as BTAN. Firstly, we incorporate propagation patterns, dispersion structure, and temporal information into a unified framework for rumor detection. Secondly, we model the temporal evolution patterns of rumors as a graph that evolves within the context of continuous-time dynamic graphs, and employ two Graph Attention Networks to capture the information regarding both top-down and bottom-up rumor propagation. Finally, extensive experiments conducted on real-world datasets show the superior performance of our model for rumor detection.</p>