Deep Chain-of-Preference: A Novel Deep Learning Method for Mixed-Grained Recommendation

The trade-off between recommending specific versus diverse information to users has long been a challenging issue in recommendation systems. In this study, we probe into a novel problem—mixed-grained recommendation (MGR)—to address this challenge. MGR involves determining the optimal recommendation granularity that aligns with users’ needs for item exploitation and exploration. To this end, we propose a novel deep chain-of-preference learning strategy to infer a user’s choice across mixed-grained categories and items, based on category-aware demand-perception alignment in a top-down manner. Specifically, we design a chain-of-preference-empowered deep learning method (CoPDL) that can infer a user’s (1) dynamic and interrelated mixed-grained demands along a multi-granularity item tree, (2) self-adapted perception along the item tree, and (3) choice regarding mixed-grained nodes in the item tree by virtue of top-down category-aware inference. Empirical evaluation results demonstrate the superior performance of CoPDL over state-of-the-art deep learning alternatives for fine-grained, coarse-grained, and mixed-grained recommendations. Further explanatory investigations render insights into how CoPDL fulfills MGR in effectively balancing the trade-off between recommendation specificity and diversity.

• 저널MIS Quarterly · 1–35
• 토픽Recommender Systems and Techniques · Information Systems
• DOI10.25300/misq/2025/19311
• 저자Gang Chen, Tan Cheng, Jianxiong Wang, Shuaiyong Xiao, Huimin Zhao