<p>Session-based recommendation (SBR) has been widely applied in many fields such as e-commerce, streaming media, social platforms, and information push, etc. How to efficiently capturing dynamic interests is one of the most prevalent research highlights in the field of SBR. Most of existing SBR models which focus on modeling on a single session, are difficult to capture user’s dynamic interests. To address this limitation, we propose an innovative graph neural network model with deformable attention mechanism on session and global graphs (DAMSGG). Specifically, the model respectively constructs session graph and global graph for fully utilizing global information. The DAM module quickly and effectively catches user’s dynamic interests and historical behaviors, while the frequency encoding module proposed effectively represents the rich information of repetitive items by calculating the occurrence counts of items. The efficiency of DAMSGG compared with various SOTA sequential models is confirmed by extensive experiments on the three public datasets.</p>