毕业生知识库

    随着互联网和大数据技术的高速发展与深度融合，互联网广告已逐步成为一种新兴的商业模式，并已成为广告主最有效的营销方式和互联网企业最重要的营收来源。在互联网广告中，关键字竞价广告和实时竞价广告是两种最为典型的互联网广告市场模态。与传统的互联网广告通过离线谈判和按时段计费售卖固定的广告位相比，关键字竞价广告实现了从"媒体购买"到"广告位购买"的模式演进，而实时竞价广告则更进一步，实现了"目标人群购买"的精准营销。

    由于关键字竞价和实时竞价广告具有低成本、高效率、强大的目标受众发现能力以及效果的易衡量性等无可比拟的优势，自诞生之日起便受到众多广告主的一致认可，并呈现出爆发式的增长态势。关键字竞价和实时竞价广告的出现极大地丰富和延展了计算广告学、互联网竞价和在线营销等研究领域的内涵，并引发了计算广告学领域新一轮的研究与应用热潮。目前，关键字竞价和实时竞价广告已成为互联网广告市场中两种最主流的商业模式和计算广告学研究中最具研究和实践价值的前沿领域。由于产生时间和背景的差异，从关键字竞价到实时竞价的商业模式演进过程必然引发新的竞价参与者及其决策场景的改变。本课题紧紧围绕这种改变，从广告需求方入手，深入挖掘广告需求方所面临的具有关键字竞价和实时竞价特色的决策问题，致力于为广告需求方提供决策支持和理论依据。

    随着参与者规模的迅速增长和竞价市场复杂性的日益提升，关键字竞价和实时竞价市场目前已呈现出前所未有的高度复杂性、动态性、强耦合性和不可预测性，这就给广告需求方的相关决策带来了极大的困难和极其严峻的挑战。一方面，在关键字竞价市场中，广告主的预算决策不仅要面临不确定环境下预算的时序分配问题，还要面临具有强耦合性的多广告计划之间的预算分配挑战。另一方面，实时竞价广告中的流量十分巨大，如何为每一个广告主选择最优匹配的流量，并最大化其广告预算的投资回报率，已成为广告需求方面临的重要决策难题。
    针对以上问题与挑战，本文基于计算实验、最优控制、随机规划、数学规划等理论与方法，研究关键字竞价广告和实时竞价广告中需求方所面临的相关决策问题，从而为广告需求方的决策提供理论依据。针对关键字竞价广告，主要研究广告主所面临的单广告计划和多广告计划预算优化问题；针对实时竞价广告，主要研究广告需求方（包括广告需求方平台DSP和广告主）的目标市场细分、广告展示频次控制和盈利模式优化等问题。

本文的主要工作包含以下几个方面：

    (1) 针对广告主在关键字竞价广告中面临的单广告计划分时段预算优化问题进行了研究。采用随机变量来刻画搜索竞价市场的不确定性，基于随机规划方法建立了单广告计划的随机分时段预算分配模型，并研究了在某些特殊情况下模型的相关性质和求解方法。利用实际推广数据设计计算实验对所提出的模型进行验证与评估，实验结果表明，考虑随机性可以使广告主获得更高的期望收益。此外，与均匀分布相比，将随机性用正态分布进行刻画时能够获得更好的广告效果。

    (2) 基于最优控制理论针对广告主在关键字竞价广告中同时推广多个具有耦合关系的广告计划时面临的预算优化问题进行了研究。首先提出广告计划重叠度的概念，从推广时间、推广地域和广告内容三个维度对具有替代关系的两个广告计划之间的关系进行描述。在此基础上，基于最优控制理论建立了多广告计划预算优化的最优控制模型，并探讨了仅有两个广告计划时模型的性质和求解方法。最后利用实际推广数据设计计算实验，对所提出的模型进行验证与评估。实验结果表明：广告计划之间的重叠度越高，广告主的最优支付越低，在预算决策中考虑广告计划之间的重叠度可以获得更高的收益。

    (3) 针对实时竞价广告中广告需求方平台（DSP）所面临的目标市场细分问题进行了研究。基于数学规划方法建立了目标市场细分模型，并给出了模型的求解算法。利用计算实验方法设计相关计算实验，对所提出的模型进行验证与评估，实验结果表明，DSP的市场细分策略对广告主的收益具有重要影响，随着市场细分粒度的逐渐增加，广告主的收益呈现出先升后降的趋势，并且DSP平台所代理的广告主数量对最优市场细分粒度的选择具有重要影响。

    (4)针对实时竞价广告中广告主面临的广告展示频次控制问题进行了研究。首先提出广告效果折现因子的概念来刻画随着同一广告向同一个目标受众展示次数的增加而产生的广告效果的变化，然后针对实时竞价广告中的两阶段竞价过程，建立了一个广告展示频次控制模型，并利用计算实验方法设计了相关计算实验来验证存在预算限制和不存在预算限制两种情况下广告展示频次对广告主收益的影响。实验结果表明，在不同的折现率下，最优的广告展示频次也不同，并且折现率越低，最优的广告展示频次越小。此外，存在一个最优的广告展示频次，使得广告主在不考虑折现率因素时能够获得最大的期望收益。

    (5) 针对目前实时竞价广告市场中现有的DSP平台的两种盈利模式（佣金模式和双层转售模式）进行了分析与探讨。首先建立了DSP和广告主在两种盈利模式下的收益模型，然后通过对DSP和广告主在两种盈利模式下的收益情况进行分析和比较，得出佣金比例取不同值时DSP和广告主各自的最优盈利模式。最后设计计算实验，验证模型及其相关性质。实验结果表明：盈利模式对于DSP和广告主双方的收益，以及佣金比例对DSP和广告主的盈利模式选择都具有重要影响。

    Online advertising has become a novel business model with the rapid development and deep integration of Internet economy and big data analysis. It has also been regarded as the most effective marketing means for the advertisers and the most important revenue source for Internet companies. At present, keyword auction and Real Time Bidding (RTB) are the most two important business models of online advertising. Comparing with the traditional online advertising which sells the fixed ad slots with off-line negotiations or prices only according to the time slots, keyword auction realized the transition from "media buying" to "ad slot buying", and furthermore, RTB advertising realized the business logic of ad-delivery from buying "ad-impressions" in websites or ad slots to directly buying the best-matched "target audiences", and thus can help advertisers achieve the precision marketing.

    Due to the advantages such as lower cost, higher efficiency, stronger ability for targeting the audiences and easier measurable, keyword auction and RTB advertising have been recognized as the most effective promotion ways for advertisers, and developed very quickly since they appeared. The appearance of keyword auction and RTB greatly enriched the connotation of the research in the field of computational advertising, Internet bidding and online marketing, and brings a new round of research and application in the field of computational advertising. At present, keyword auction and RTB advertising have become the most two popular business models in Internet advertising, and also with the most research and practical value in computational advertising. Due to the differences of their emerging time and backgrounds, the evolution of the business model from keyword auction to RTB advertising will inevitably lead to the changes of new bidding participants and decision making scenario. Based on such changes, this research focused on the advertising demand sides, and studied their decision issues in keyword auction and RTB, aiming to provide some decision supports and theoretical basis for the advertising demand sides.

    With the rapid growth of the participants and the increasing complexity of the markets, keyword auction and RTB advertising markets have shown an unprecedented high complexity, dynamic, strong coupling and unpredictability, which bring great difficulties and challenges to the decision-making processes for the demand sides. On one hand, the difficulties in the budget allocation decisions for the advertisers in keyword auctions are mainly induced by the uncertain market environments and the coupled relationships among different campaigns. On the other hand, there are vast amount of ad requests in RTB market, and how to choose the best-matched ad requests and maximize the advertising effect for the advertisers have become the most important and difficult decision problems for Demand Side Platforms (DSPs) and advertisers in RTB advertising.

    The purpose of this work is to deal with the above difficulties and challenges faced by the demand side in keyword auction and RTB advertising, and provide some necessary theoretical foundation for them. By utilizing computational experiments approach, optimal control theory, stochastic programming and mathematical programming approach, we study the budget optimization problem faced by the advertisers with one campaign or multiple campaigns in keyword auction, and the issues including market segmentation, frequency capping and the revenue models for the demand sides (i.e., DSP and advertisers) of RTB markets. The main researches and contributions of this work include the following five aspects:

    (1) We studied the budget optimization problem over a series of sequential temporal slots faced by the advertisers in keyword auctions when only one campaign was promoted. Since there exists large uncertainty in search markets, we introduced random variable to characterize it, and presented a stochastic model for budget distribution over a series of sequential temporal slots under a finite time horizon. Some properties and possible solutions for the proposed budget distribution model were studied when the random variable was characterized by uniform and normal distributions. We also conducted some computational experiments to evaluate the proposed model, and normative findings from the experiments show that, the expected revenues of the advertisers can be greatly improved by considering the randomness in their budget planning strategies, and the performance of normal distribution is better than uniform distribution.

    (2) Budget planning over several coupled campaigns in keyword auctions was studied based on the optimal control theory. Firstly, a three dimensional measure of substitution relationships between campaigns is presented, namely the overlapping degree in terms of promotional periods, target regions and campaign contents. After that, we proposed a dynamic multi-campaign budget planning approach using optimal control theories, with consideration of the substitution relationship between advertising campaigns, and some desirable properties of our budget model and possible solutions were also studied for the case that there are only two campaigns. Finally, computational experiments were conducted to evaluate our model using real-world data of practical campaigns in keyword auctions, and experimental results showed that, coupled campaigns with higher overlapping degree can reduce the optimal payoff, and the advertising effort can be seriously weakened by ignoring the overlapping degree between campaigns.

    (3) We studied the market segmentation issue faced by the DSPs in RTB advertising. Based on a mathematical programming approach, we established an optimization model of the granularity for segmenting RTB advertising markets, and its solution algorithm was also provided. We proposed to validate our model using the computational experiment approach, and the experimental results show that, the market segmentation granularity has the potential of improving the total revenue of the advertisers, and with the increasing refinement of the market segmentation granularity, the total revenue has a tendency of a rise first and followed by a decline. Moreover, the optimal granularity of market segmentation will be significantly influenced by the number of advertisers on the DSP.

    (4) We studied the frequency capping problem faced by advertisers in RTB advertising. Firstly, the discount rate was introduced to characterize the diminishing of the advertising effect when displaying the advertisements multiple number of times to the same target audience. After that, we established an optimization model for frequency capping considering the two-stage bidding process in RTB advertising. Finally, utilizing the computational experiment approach, we designed two experiments to validate the proposed model, and the experimental results show that under different discount rates, the optimal frequency caps are different. Moreover, the optimal frequency cap is smaller for lower discount rate, and  when considering all the discount rates, there exists an optimal frequency cap, at which the expected maximum revenue can be obtained in the long run.

    (5) We studied the existed two revenue models for DSPs in RTB advertising, namely the commission model and the two-stage resale model, respectively. We first proposed the methods for computing the revenues for DSPs and the advertisers under these two models, and then compared their revenues in different cases. By comparing their revenues, we deduce the optimal revenue model for DSPs and advertisers under different commission rates. We also design computational experiments to evaluate our proposed models and their properties. The experimental results show that different revenue models can bring different revenues for DSPs and advertisers, and the ratio of the commission from advertisers can greatly affect their choices of the revenue models.