王擎 wangqing.linus@gmail.com wangqing09@baidu.com Hi: wangqing09hi 2014-4-1 RTB 简介 • Real-Time-Bidding 实时竞价购买广告的方式 • AdExchange 广告交易市场 • DSP(Demand Side Platform) 需求方平台,为广告主 的推广campaign优化投放 • SSP(Supply Side Platform) 供应方平台,为媒体(广 告位)进行服务管理 • DMP(Data Manage Platform) 数据管理平台,管理 不同来源的数据 RTB 简介 Picture from: http://contest.ipinyou.com/manual.shtml RTB 简介 Picture from: http://contest.ipinyou.com/manual.shtml RTB 特点 • 建立统一市场环境,角色细分 • (对于广告主) 网络媒体包买广告位 -> 每个曝光单独购买 • (对于媒体网站)提供广告位&观众的 交易平台,有效利用资源 advertisement audience iPinYou • DSP: • 人群分类(Audience Targeting) – Cookie(1x1 pixel), AdEx Cookie Mapping, ... – > User Tags • 实时交易(RTB) – 基础架构 ~100ms – Bidding Model iPinYou 拍卖规则 • 二价模型(Vickrey Auction or VCG) • Highest price wins, pays 2nd-highest price. • 鼓励真实出价 "The dominant strategy in a Vickrey auction with a single, indivisible item is for each bidder to bid their true value of the item." cf. wikipedia: Vickrey_auction 模型目标 Maximize Nclicks + N * Nconversions subject to the fixed ad budget. budget -> impression -> click/conv CPI, CTR -> CPC 比赛流程 • 3个赛季 • 第二赛季: 1.线下leaderboard (6月-9月) 选出前5名 2.上线前调试 3.正式比赛3天 (9月中旬) 最终排名 • 3个Ader • 4个AdExchange Google Ali tencent Baidu -> -> 赛季结果 score = clicks + N * reaches Team Score N=20 Team CAS_MLRush 8429 o_o 4289 UCL-CA 1304 the9thbit 8229 梦想照进现实 4067 V_V 983 newline-CA 6844 UCL-CA 1995 Run Fast 901 Again 1409 PoundsXXX 885 deep_ml 1148 Tiger 744 http://contest.ipinyou.com/message_list.shtml Score N=6 Team Score N=1 Our Approach 1. Feature Extraction 2. CTR Prediction 3. Online Bidding Features -> pCTR -> bid_price Feature Extraction • low-level + high-level • CreativeId, AderId, AdExchange(Platform), Domain, AdslotId, AdslotWidth, AdslotHeight, AdslotVisibility, AdslotFormat, Floorprice, UserTags, OS, IE(Browser), Region, City, Weekday, Hour Feature Extraction adslot user ad pCTR: Logistic Regression Algorithm J.Langford, L.Li, T.Zhang, Sparse Online Learning via Truncated Gradient Vowel Wabbit https://github.com/JohnLangford/vowpal_wabbit/ Also cf. C.Perlich et al. Machine learning for targeted display advertising: Transfer learning in action (SGD + elastic net) Algorithm Bidding Bid price: β 约为 2.0 .. 3.0 Also cf. Season1江申的分享, 及Media6Degrees : Bid Optimizing and Inventory Scoring in Targeted Online Advertising (Step function) “Any ratio below 0.8 yields a bid price of 0 (so not bidding), ratios between 0.8 and 1.2 are set to 1 and ratios above 1.2 bid twice the base price.” Offline results Offline results Online results (Sep.16, Ader 2821) Online results (Sep.16, Ader 2940) Online results (Sep.16, Ader 3430) 遇到的挑战 • 预估CTR 1. 样本不平衡([LR in rare events]) 2. 特征稀疏(reg, naive bayes) 3. 算法的收敛(2nd order / 1st order) • 控制花钱速度 1. 流量预测 (at different bid levels) 2. 市场竞争环境建模(Bid landscape forecasting?) 3. 自动调节 预算控制 K.Lee, A.Jalali, A.Dasdan, Real Time Bid Optimization with Smooth Budget Delivery in Online Advertising 实际系统 • 算法: 1. 人群分类(user tags) 2. 广告预选 (rank) 3. 在线反馈 (online learning) 4. 冷启动/热启动 (transfer learning) • 架构: 1. 实时响应 并发 2. 数据日志整合 3. 其他 总结与收获 1. Feature Extraction low-level + high-level 2. CTR Prediction logistic regression 3. Online Bidding power scheme • 学习的过程 • 实践的机会 • 交流的平台 Thanks! 参考 • 比赛官网 http://contest.ipinyou.com/ • 数据下载 http://pan.baidu.com/s/1kTkGUQN • J.Langford, L.Li, T.Zhang, Sparse Online Learning via Truncated Gradient • C.Perlich, B.Dalessandro, R.Hook, O.Stitelman, T.Raeder, F.Provost, Bid Optimizing and Inventory Scoring in Targeted Online Advertising • K.Lee, A.Jalali, A.Dasdan, Real Time Bid Optimization with Smooth Budget Delivery in Online Advertising • C. Perlich, B. Dalessandro, O. Stitelman, T. Raeder, F. Provost, Machine learning for targeted display advertising: Transfer learning in action 参考 • Y.Cui, R.Zhang, W.Li, J.Mao, Bid Landscape Forecasting in Online Ad Exchange Marketplace • G.King, Logistic Regression in Rare Events • 江申的分享 http://www.techinads.com/archives/41 http://pan.baidu.com/share/link?shareid=322913515&uk =3138366223 • 品友有关DSP的介绍 • 艾瑞咨询2013年中国DSP行业发展报告