Bundle recommendation aims to provide personalized product bundles, yet existing methods suffer from redundant learning (overlapping user-item and user-bundle interactions) and undifferentiated interaction modeling (uniform aggregation of nodes). To address these challenges, we propose a novel method named Multi-interaction Graph attention network for Bundle Recommendation (MGBR), which integrates multi-interaction separation, topology-aware structural learning, and cross-task preference transfer. Specifically, MGBR constructs three homogeneous graphs (user-bundle, user-item, bundle-item) to isolate interaction redundancy and employs a hierarchical graph attention mechanism to dynamically assign weights to interactions (e.g., distinguishing core vs. peripheral items in bundles). To enhance structural discriminability, we propose a structural hint learning module with dual objectives: (1) node degree prediction to preserve user activity and bundle popularity patterns; (2) neighbor degree sum prediction to capture local topological dependencies. Additionally, multi-task learning transfers knowledge between item-level and bundle-level preferences through shared user embeddings. Extensive experiments on two real public datasets demonstrate that MGBR outperforms previous bundle recommendation methods by 1.30%–4.65% on NetEase and Youshu.