The Next Basket Recommendation (NBR) task focuses on predicting a set of items for users based on their interests. Although utilizing preference information from users’ historical sequences can enhance NBR accuracy by capturing general preferences, these methods struggle when historical data is sparse. Additionally, modeling customized preferences for different users remains a challenge. In this work, we introduce a novel Collaborative Aggregation Model for NBR with Time-independent Sequence Modeling, namely CoANBR, which comprises three modules: (1) The Item-aware Aggregator constructs an item-item graph using user-item and set-item interactions, leveraging graph learning to obtain item embeddings and effectively aggregate neighborhood information. (2) The User Preference Enhancer captures interactions among items to help users discover potential preferred items globally, and further enhances general preferences to address data sparsity issues. (3) The Decoupled Temporal Attention Module models the temporal dependencies of item positional information within interaction sequences, enabling the customization of user preferences across different contexts. Extensive experiments on real-world datasets show CoANBR’s effectiveness, consistently outperforming several state-of-the-art NBR methods.