实时竞价在展示广告中的应用研究及进展  

Research Progress of Real-Time Bidding for Display Advertising

作　　者：刘梦娟[1] 岳威 仇笠舟 李家兴 秦志光[1] LIU Meng-Juan;YUE Wei;QIU Li-Zhou;LI Jia-Xing;QIN Zhi-Guang(Network and Data Security Key Laboratory of Sichuan Province,University of Electronic Science and Technology of China,Chengdu 610054)

机构地区：[1]电子科技大学网络与数据安全四川省重点实验室,成都610054

出　　处：《计算机学报》2020年第10期1810-1841,共32页Chinese Journal of Computers

基　　金：国家自然科学基金(61202445,61502087);中央高校基本业务费(ZYGX2016J096)资助.

摘　　要：随着在线广告在产业界取得巨大成功,其在学术界特别是数据挖掘和机器学习领域的研究也吸引了大量学者的关注.本论文围绕实时竞价机制在展示广告投放中的关键问题展开研究.首先介绍了实时竞价的基本流程、主要参与者的功能、定价模型和交易机制;然后分别从需求方、供应方和交易中心的角度,介绍了实时竞价中存在的关键问题,以及目前的研究方法、理论和模型,具体包括:用户响应预测、出价策略、预算与步进管理、保留价优化、库存分配、拍卖机制等,特别针对用户响应预测和出价策略两个研究热点展开了广泛讨论,并对其中的代表性方法进行了量化对比;此基础上对主要的广告欺诈方式和检测手段进行了整理;最后对该方向未来的研究趋势进行展望.During the past several years,online advertising has achieved huge success in the industry.As one of the most exciting advances in online advertising,real-time bidding(RTB)has attracted a lot of attentions from both academia and industry,especially in the fields of data mining and machine learning.Different from the traditional contract-based ad delivery,in RTB the publishers(such as websites and mobile apps)can sell their ad impressions through public auctions and advertisers can evaluate the auctioned impressions and bid for them.Therefore,in RTB publishers can sell more impressions and make more money and advertisers can optimize their budget efficiency by allocating the budget to all of the available impressions based on their values.This paper focuses on the research progress of RTB for display advertising.Firstly,we briefly introduce the ad delivery process,the functions of main participants,the pricing models and auction mechanisms in RTB.Secondly,from the perspectives of demand side platform(DSP),supply side platform(SSP)and ad exchange(ADX),we detail the key problems in RTB,as well as the existing research methods,theories and models.Specially,DSP plays a critical role as the agent for advertisers,where we introduce three topics,i.e.user response prediction,bidding strategy and budget(pacing)management.Among them the first two topics are the most popular research areas.For user response prediction,there are three structures that are widely used in prediction models:the first one is shallow structure,the second one is deep neural network,and the last one is the hybrid structure.We introduce the representative models based on the above structures and evaluate their prediction performance based on a real-world dataset.Bidding strategy is used to decide the bid price for an impression on behalf of the advertiser.In this paper,we introduce the widely used linear methods and the latest nonlinear and reinforcement learning based methods.Empirically,we compare the characteristics and performance of some typical

关 键 词：展示广告 实时竞价 点击率预测 出价策略 广义第二价格拍卖 广告欺诈 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]