Intra-Category Multi-Choice Preferences Learning and Assortment Recommendation in E-Commerce

• 成果类型: Article; Early Access
• 署名作者: Lin, Hongyuan; Li, Xiaobo; Wu, Lixia
• 署名单位: Chinese Academy of Sciences; University of Science & Technology of China, CAS; National University of Singapore
• 刊物名称: PRODUCTION AND OPERATIONS MANAGEMENT
• ISSN/ISSBN: 1059-1478
• DOI: 10.1177/10591478251350853
• 发表日期: 2025
• 关键词: E-commerce; Assortment Optimization; Multi-purchase behavior; Non-parametric choice model; Click-stream data

摘要

This paper proposes a framework for learning customer preferences and optimizing e-commerce assortments, focusing on intra-category multi-choice behavior, where customers buy multiple items within the same category. Unlike traditional discrete choice models (DCMs) that assume single-product purchases, e-commerce data show frequent intra-category multiproduct purchases, especially during promotions. To capture this, we introduce the multi-choice rank list model (MC-RLM), which accounts for both multi-purchase and substitution effects. Each customer type is defined by a preference ranking and an intended purchase quantity (IPQ), allowing selection of up to IPQ products. The MC-RLM adheres to the regularity axiom and aligns with random utility theory. We present the multi-choice market discovery algorithm to estimate the MC-RLM, extending single-purchase methods to multi-purchase settings. We also introduce behavior-reveal-preference (BRP) rules, using customer behavior data (e.g., clicks, cart additions) to enhance preference estimation. Given the NP-hardness of the assortment optimization problem, we analyze the performance of the revenue-ordered assortments heuristic and provide guarantees. The problem is formulated as a mixed-integer linear program that can generate personalized recommendations based on real-time customer data. Extensive numerical experiments, including a case study using Tmall data, demonstrate that the MC-RLM outperforms models such as the independent choice model and the multi-purchase multinomial logit model in predictive accuracy, with BRP rules further enhancing performance. Synthetic experiments confirm that accurately modeling multi-purchase behavior significantly boosts expected revenue.