No1.Simple and Scalable response prediction for display advertising
点进去和回话率是展示广告中两个核心预测指标。
关键字：特征选择，分布式学习，点击预测，散列法
1.1 What is Display advertising?
1.2 Spot markets offer wide range of payment.CTR is cirtical.
1.3 Signicant work in search market for modeling.
1.4 Provide a machine learning framework. Use Logistic or Maximum Entropy.
1.5 Experimental results show their framework outperfoms.
1.6 Part 2 related work
      Part 3 difference of CTR and convesation rate
      Part 4 Maximum Entropy
      Part 5 result
      Part 6 modied model and experimentle result
      Part 7 a algorithm for exploration
      Part 8 implementation of map-reduce
2.1 related work classification:
       content match context features   
       text ads and display ads
       special features
       feedback features from history
2.2  some features is unavailable
2.3  Logistic regression and decisiontree
       LR is easy for handling large scale problems
2.4  To use sparse data to predict-Possion
       why less domain knowledge is used but useful.
3.1  Data from Right Media Exchange
        ten billion ad impressions one day
        site id ,tag id,pricing type,budget, targeting prole
3.2   post-click conversion PCC click through rate CTR conversion rate CVR
3.3   information sources needed for PCC CTR CVR
       Advertiser: advertiser id ,advertiser network,campain,creative,conversion id,ad group,ad size,creative type,
                        offer type id(ad category)
       Publisher: publisher (id),publisher network,site,section,url,page reffer
       User:gender,age,region,network speed, accept cookies,geo
       Time: serve time,click time
3.4  click conversion delay
       建模第一步：  如何把会话归于点击。会话可以是几分钟，几小时，几天。点击与会话之间的延时也有可能是几天，几小时，几秒钟。可以通过匹配它们的特征实现。
       写下贝叶斯模型：
       \[\begin{alignat}{1}P(Click,Convertion)&=P(Click)P(Conversion|Click)\\&=P(Convertion)P(Click|Conversion)\end{alignat}\]
      考虑延时：
       \[\begin{alignat}{1}P(Delay，Click，Conversion)&=P(Delay)P(Click,Conversion|Delay)\\&=P(Click,Conversion)P(Delay|Click,Conversion)\end{alignat}\]
       注：考虑了四个随机变量：点击不点击，点击几次，延时多长，会话多长
              对应了四个概率模型：伯努利分布，Possion分布，指数分布，Gamma分布
              因为：很低的点击率，等待第一次会话发生的时间，等待第n次会话发生的时间
3.5 根据基本统计数据选训练数据集。
       延时之后的会话：
       10分钟 86.7%
        1分钟  39.2%
        1小时  95.5%
         2天     98.5%
Problems:
What is Thompson sampling?
What is baseline Probprobility?
What is matrix factorization?
What is Tkikhonov regularization?
What is conjuction features?
What is Covariates?
What is the criticle aspect lead to action?
如何将连续随机变量离散化，假设离散后的块服从什么分布，连续的情形服从什么分布，离散化之后服从什么分布？
What is the difference of sponsored search and paid search
百度搜什么推什么的策略有效吗？百度的旋转框更像个游戏会吸引人的注意力，但是我自己坚决不会去点，而且放大字框的搜索广告十分令人讨厌。
What is the relationship of CPC and bid
Decision:
1.不将Logistic，讲是与否
2.模型描述等待用户点击并且会话，key
------------------------------------------------------------------------------------------------------------------------------------------------------------
NO.2 Predicting Clicks: Estimating the Click Through Rate for New Ads ***** before 2004

Abstrct:To choose Ads to satisfy user and earn more revenue

1.1 next to their search result
1.2 different ways of payment of different interaction of ad and audience:impression,click,action,pay
1.3 Aspects that will influent Click in a ad
2.1 search engine for more ad
因为：revenue=p(click)*CPC
2.2 本文重点在基于搜所引擎竞价的工作
2.3 CTR for sort ads.The assumption is 随时间推移特定广告的CTR收敛于常值，不考虑exhibit periodic or inconsistent behavior
.CTR 确实相对较低，估计的变数相对较高，即便展示量不很多也不稀疏。比如，一个真实CTR有5%的广告必须在展示1000次之前，它的CTR 也有可能有能让我们置信85%的1%.一般的搜索广告点击率大概在2.6%左右。
2.4系统的收敛意味着大规模的所有货币话化。比如，CPC1.6需要50次80块的点击进入，任何点击率预测的食物都会导致接下去的优化排序和搜索引擎丢失营收以及高表现广告低浏览量。
2.5搜索广告市场近年成长迅速，每天都有新广告主进入，同时有广告主频繁登入新广告campain。许多广告主都在创造新的campain，另外还有一些会抱尝试的目的来优化他们的广告效果。所有这些时间都会导致不断增加的需要排序的查询。
2.6另外，有的广告有时候就是在求新查询。游戏广告主试图通过猫定数以千计的鲜有搜索条目是他的每个query都被排序来优化他们的投资回报率。这在PPCcampain中有十分显著的增长。
2.7调查显示
3.1搜索广告的框架
利用点击了必然是看见了，可以做出一个贝叶斯模型；
-------------------------------------------------------------------------------------------------------------------------------------------------------------
NO 3.Predicting Response in Mobile Advertising with Hierarchical Importance-Aware Factorization Machine 2014
Primary Degree by students
NO 4.Estimation Conversion Rate in Display Advertising from Past Performance Data
Primary Degree by students
NO 5. CTR Predictions and Literature Reference
Primary Degree by studens