报告题目：The Power of Linear Programming in Sponsored Listings Ranking: Evidence from a Large-Scale Field Experiment

报告人：朱玉婷 助理教授

邀请人：叶鑫 教授

报告时间及地点：2026年5月21日 15：00-16：30 经济管理学院B305

报告人简介：

报告内容摘要：

Sponsored product advertisements constitute a major revenue source for online marketplaces such as Amazon, Walmart, and Alibaba. A key operational challenge in these systems lies in the Sponsored Listings Ranking (SLR) problem, that is, determining which items to include and how to rank them to balance short-term revenue with long-term relevance and user experience. Industry practice predominantly relies on score-based algorithms, which construct heuristic composite scores to rank items efficiently within strict real-time latency constraints. However, such methods offer limited control over objective trade-offs and cannot readily accommodate additional operational constraints. We propose and evaluate a Linear Programming (LP)-based algorithm as a principled alternative to score-based approaches. We first formulate the SLR problem as a constrained mixed integer programming (MIP) model and develop a dual-based algorithm that approximately solves its LP relaxation within 0.1 second, satisfying production-level latency requirements. In collaboration with a leading online marketplace, we conduct a 19-day field experiment encompassing approximately 329 million impressions. The LP-based algorithm significantly outperforms the industry-standard benchmark in key marketplace metrics, demonstrating both higher revenue and maintained relevance. Mechanism analyses reveal that the performance gains are most pronounced when the revenue-relevance tradeoff is stronger. Our framework also generalizes to settings with inventory, sales, or fairness constraints, offering a flexible and deployable optimization paradigm. The LP-based algorithm was deployed in production at our partner marketplace in January 2023, marking a rare large-scale implementation of a mathematically grounded ranking algorithm in real-world online advertising.

译文：

赞助商品广告是在线市场平台（如亚马逊、沃尔玛和阿里巴巴）的主要收入来源之一。这类系统中的一个关键运营挑战是赞助列表排序（Sponsored Listings Ranking, SLR）问题，即如何决定展示哪些商品以及如何对其进行排序，以在短期收入与长期相关性和用户体验之间取得平衡。当前行业实践主要依赖基于评分（score-based）的算法，通过构建启发式的综合评分对商品进行排序，从而在严格的实时延迟约束下实现高效决策。然而，这类方法对不同目标之间的权衡缺乏精细控制，也难以灵活纳入额外的运营约束。本文提出并评估了一种基于线性规划（Linear Programming, LP）的算法，作为基于评分方法的一种原理性替代方案。我们首先将 SLR 问题形式化为一个带约束的混合整数规划（Mixed Integer Programming, MIP）模型，并开发了一种基于对偶的算法，能够在 0.1 秒内近似求解其线性规划松弛（LP relaxation），从而满足生产系统对实时延迟的要求。在与一家领先在线市场平台合作的背景下，我们开展了一项为期 19 天的现场实验，覆盖约 3.29 亿次展示（impressions）。实验结果表明，与行业标准基准算法相比，基于 LP 的算法在关键市场指标上表现显著更优，同时实现了更高的收入并保持了广告相关性。机制分析进一步发现，当收入与相关性之间的权衡更为显著时，该算法带来的性能提升最为明显。此外，我们提出的框架还能够推广到包含库存约束、销售约束或公平性约束的场景，为在线广告排序提供了一种灵活且可部署的优化范式。LP 排序算法已于 2023 年 1 月在合作平台的生产系统中正式上线，成为现实在线广告系统中少见的大规模数学优化驱动排序算法的落地应用案例。