数据驱动的在线广告定向投放决策研究

Targeted delivery gives online advertising a competitive edge, it allows advisers to deliver their ads to preferred online customers, significantly strengthening the effectiveness of advertising. It is a central decision problem for ad publisher that how to perform targeted ad delivery. Through the investigation of some leading online advertising firms in the industry, we have found that the following four issues exist in the current practice: (1) inaccurate estimation of page views makes it hard to complete delivery tasks for some guaranteed ads; (2) the actual performance of auction mechanism is lower than theoretical expectation; (3) the diminishing marginal effect of ad exposure has been overlooked; and (4) ad publisher can miss the profit potential because of separate management of guaranteed ads in the upfront market and non-guaranteed ads in the spot market...To address the above issues, this project proposes an integrated planning for targeted ad delivery, holistically taking in to account the sharing and competition on page views between guaranteed ads and non-guaranteed ads; and proposes to replace the auction by planning approach. On basis of real data, this project forecast ad clicks, build up and solve a robust optimization model for online ad targeted delivery that maximizes ad publisher’s revenue, while ensuring the fulfillment of delivery task and effectiveness under the uncertainty of page views, achieving a revenue-efficiency balanced ad delivery. By utilizing real data, this project proposes to carry out extensive ad delivery experiments in real environment, evaluating and improving research outputs. Our ultimate goal is to fully support ad publishers from managerial decisions to practical applications. Consequently, this project has substantial scientific significance and applicable value.

定向投放是在线广告的核心竞争力，它使得广告主向特定用户投放广告，极大增强了广告效果。如何进行定向投放是广告发布商重要的决策问题。在行业领先的企业调研中，我们发现当前实践存在四个弊端：(1)流量预估值不准确，导致部分品牌广告投放任务难以完成；(2)竞价拍卖机制实际表现低于预期；(3)忽视广告投放效果边际递减效应；(4)独立管理品牌广告与竞价广告，错失潜在收益。..为解决上述实践弊端，本项目拟提出品牌广告与竞价广告的整合式定向投放思想，并利用规划投放取代竞价拍卖。本项目基于实际数据，预测广告点击次数，建立并求解在线广告定向投放鲁棒优化模型，最大化广告发布商的收益；同时在不确定的流量下，保证定向投放的任务与效果，实现效益均衡的投放。另外，本项目拟利用实际数据，进行真实环境下的广告投放实验，评估与改进研究成果，为广告发布商提供从管理决策到实际应用的全面支持。因此，本项目具有重要科学意义与应用意义。

我国在线广告行业近十年来得到了迅猛发展，其市场规模已远超传统广播和电视广告。在线广告提供了精准的定向投放，使得广告主能向特定的目标用户投放广告，极大增强了广告营销效果。在数字技术赋能企业决策的背景下，本项目从广告发布商的角度出发，研究了基于数据驱动的在线广告定向投放决策。在过去的三年里，项目团队按计划展开工作，取得了一些创新型的研究成果。（1）针对竞价广告收益管理，提出用规划投放替代竞价拍卖，解决了竞价拍卖容易错失潜在收益的不足。设计了一个由广告点击次数预测和广告投放优化组成的数据驱动广告投放技术框架，其中，预测部分考虑了用户点击广告行为和广告投放效果的边际递减效应，优化部分为广告发布商建立了一个混合整数非线性规划模型以最大化收益，并设计了高效的求解算法。（2）针对同时运营品牌广告和竞价广告的发布商，提出了一种整合式的广告投放策略，综合考虑两种广告的资源共享与竞争。具体而言，在不确定的流量下，为整合式广告投放建立了一个分布式鲁棒概率约束优化模型，广告发布商可以灵活调整风险偏好，平衡品牌广告的投放效果和竞价广告的投放收益；将此模型最终转换为线性规划模型，证明了转换有效性和转换带来的目标函数值损失的理论上界，并针对大规模实例设计了高效求解方法。项目负责人将相关研究成果以第一作者身份分别发表于管理科学领域国际顶尖期刊Operations Research和Production and Operations Management。项目研究成果丰富了在线广告定向投放的理论与方法，促进了数据驱动型预测与优化方法在企业实际决策中的应用。