The session-based recommendation predicts a user’s next action based on their behaviors within the current session. Unlike traditional recommender systems that rely on long-term user data, this system treats users anonymously, without historical information. The main challenge is to extract relevant user preferences from the session while minimizing noise. To address these challenges, we propose Multi-granularity User Intent and Dual-channel Sparse Graph Attention Networks (MG-DSGAT). This model employs a dual-channel sparse attention network to capture both inner-session and cross-session information, enhancing accuracy and reducing the impact of irrelevant items. An additional attention network extracts multi-granularity user information, and a fusion module integrates inner-session, cross-session, and multi-granularity intent to better model complex user behaviors. Extensive experiments on three real-world datasets demonstrate that MG-DSGAT outperforms state-of-the-art session-based recommendation models.