Multi-scale Context-aware User Interest Learning for Behavior Pattern Modeling

Next Basket Recommendation (NBR) mines user interests from sequential basket records where multiple items are purchased together. Existing methods face two challenges: 1) insufficient modeling of behavioral patterns, leading to coarse-grained user interest learning; and 2) information loss in user interest learning, leading to suboptimal results. We propose a novel solution, Multi-scale Context-aware Recrecommendation (MCRec), which overcomes these issues by mapping basket sequences to tensors for latent space representation learning. MCRec employs vertical, horizontal, and dilated convolutions to extract multi-scale context-aware user interests that capture diverse behavioral patterns. Specifically, MCRec integrates an adaptive user interest fusion mechanism for multi-level user interest modeling, which combines user representations from historical records with preferences derived from interaction frequencies for accurate predictions. Extensive experiments on three real-world datasets demonstrate that MCRec outperforms several representative NBR methods and achieves state-of-the-art results.