Speaker: 高品，香港中文大学（深圳）助理教授

Time/Date: 2024月5月22日 10：30

Classroom: 同济大厦A楼408教室

ABSTRACT:

This paper introduces a novel regular choice model that merely imposes an upper bound on each product’s last-choice probability (i.e., the purchase probability of being exclusively recommended). Notable instances of this model include the Multinomial Logit (MNL), click-based MNL, sequential click-based MNL, and the general attraction model. For the static cardinality-constrained revenue-maximizing recommendation problem, despite its NP-hardness even under specific model instances, we propose a polynomial-time solvable constant-factor approximation heuristic for the general case, leveraging solely the last-choice probabilities. Additionally, we investigate an online recommendation scenario, in which product-specific parameters are sequentially revealed as random and independent variables, compelling decision-makers to make immediate and irreversible decisions regarding whether to recommend a product upon receiving its information. We offer constant-factor approximations for cases where the information arrival order is adversarial or adheres to specific distributional patterns. Our comparative performance analysis underscores the importance of information in the online setting and establishes a novel connection between the online recommendation problem and the classical multiple-unit prophet inequality. 本研究深入探讨了一般选择模型框架下的品类优化问题，并提出将复杂问题简化为独立需求模型下的问题。在简化模型中，每个产品的收益将减去一个相同的常数，以弥补未考虑替代效应所带来的损失。结果表明，简化模型下所获得的启发式算法在原模型下所产生的收入至少占最佳收入的一个固定比例，而该比例取决于原模型的两个关键参数。当这两个模型参数确定时，我们的表现比率是最优可获得值。在模型参数难以计算的情况下，我们提出了一种有效的算法，能够任意近似达到最佳的表现比率。

GUEST BIO:

高品博士是香港中文大学（深圳）数据科学学院的助理教授。他本科毕业于武汉大学物理系，并在香港科技大学获得物理学硕士和运筹学博士学位。高博士的研究兴趣包括收益管理和新兴商业模式中的运营管理。他有多篇论文已在Management Science，Operations Research等顶级期刊上发表。高博士获得多个研究奖项，包括ISCOM最佳论文奖（第一名）和POMS-China最佳论文奖（第二名）等。