机器学习在互联网广告中的应用 庄宝童
Agenda 介绍 机器学习应用 Common utility Advertiser Publisher user 总结
为什么需要互联网广告? 流量(用户)是互联网公司的重要资产 互联网内容免费模式,需要流量变现来维持运营 广告收入占比: Google :95% (2012,http://investor.google.com/financial/tables.html) Facebook:83% (2011) Baidu:? Alibaba:? 特点:效果量化可追踪,运营销售参与少,曝光成本低 对互联网广告公司而言,是一种理想的“印钞机”商业模式(吴军,《浪潮之巅》)
我们需要什么样的广告? Find the best match between a given user in a given context and a suitable advertisement -- Andrei Broder and Dr. Vanja 2011 各种utility: GSP CPA CPC CPM
Advertisers Ads Page Ad Network User Publisher Pick best ads Response rates (click, conversion, ad-view) conversion Bids Auction Statistical model Advertisers Ad Network Ads Page Pick best ads User Publisher Select argmax f(bid, rate)
Players in the ecosystem Publisher’s utility:Revenue,user engagement Advertiser ‘s utility:ROI User’s utility:relevance
mechanism design 合同定价 ( futures market),CPM 或 CPT 计价 拍卖定价 (spot market) GFP GSP VCG 计价方式 CPM (Cost per Mille-impressions): publisher 风险最小,如 yahoo,sina的品牌广告 CPC (Cost per Click) : publisher 和 advertiser 风险共担,google adwords,百度凤巢等大部分属于此类 CPA (cost per Action):advertiser 风险最小,如淘宝客。
CPC 的ranking functions Bid ranking:bid 源于 goto.com (overture 前身,后被yahoo收购) Revenue ranking:CTR * bid Google 首创 核心问题:CTR prediction
model P(click | user, ad, context) ad : creative, bid-terms, landing page, campaign, advertiser, format (text/image/video), size, etc. user : cookie, demo, geo, behavioral, activity history context : query, publisher, page-content, session, time
algorithms Logistic Regression + feature engineering (google, yahoo, baidu, facebook , etc) Microsoft (Baysian Probit Regression) Google : boosting http://users.soe.ucsc.edu/~niejiazhong/slides/chandra.pdf Taobao (Mixture of Logistic Regression) trends:big data + nonlinear/feature learning
challenges Sparsity: use Natural hierarchies or Auto-generated hierarchies Missing data Bias:position,ad category,etc Dynamical /seasonal effects Spam/noisy data
features Features: Preprocess: Click feedback features (COEC) Query features Query-ad text matching features Preprocess: 离散化 分段 特征交叉 层次特征—处理稀疏性 (variance bias trade-off) 特征平滑,变换
training 训练集 分布式训练 MPI (baidu, taobao) map reduce (google) 正负样本分层采样 – imbalance training 问题 Instances:1B Features:10B 分布式训练 MPI (baidu, taobao) map reduce (google)
Evaluation Offline evaluation Online A/B test MSE, MAE AUC 分层实验平台(google,Overlapping Experiment Infrastructure: More, Better, Faster Experimentation) 正态/二项分布样本的假设检验
实践 实时计算,性能问题 简单有效的候选集选取 精确计算 Online learning
Explore/Exploit 低 mean ,高方差的 ads 应该給予展示机会 E.g. Consider 2 ads (same bids) Goal: Select most popular CTR1 ~ (mean=.01,var=.1), CTR2~ (mean=.05,var~0) CTR Probability density Ad 2 Ad 1
E&E 常用算法 Upper confidence bound policy (UCB) Thompson sampling 问题 Mean + uncertainty-estimate mean + k* sd(estimator) Thompson sampling 从 posterior 里随机采样,比较适合 Bayesian 类的算法 问题 广告集合巨大,explore 代价过大 跟传统 Multi-Arms bandits 问题不太一样,广告集合是动态的,且每次会选择多个
Advertiser’s perspective Keyword selection Bid optimization Smart pricing Anti fraud Impression forecasting: time series Smooth delivery: allocation algorithms
CVR prediction 用途: 做法:与CTR 预估问题类似,但更困难 Smart pricing :外部流量千差万别,广告主没有精力也能力做分媒体的出价,需要按照点击价值进行智能出价 (Google, smart pricing grows the pie),以保证广告主的 ROI DSP: real time bidding CPA 模式的rank function: ctr * cvr * bid 做法:与CTR 预估问题类似,但更困难 转化数据获取困难,且更为稀疏 不同广告主的转化定义不一致
User’s perspective User fatigue User privacy Behavioral targeting / retargeting Query intent Low quality ads detection(google, detecting adversarial advertisements in the wild)
Publisher’s perspective Revenue User engagement
谢谢