【RTB论文笔记】

Bidding Strategies

Attribution Modeling Increases Efficiency of Bidding in Display Advertising by Eustache Diemert et al. ADKDD 2017.
Profit Maximization for Online Advertising Demand-Side Platforms by Paul Grigas et al. ArXiv 2017.
Real-Time Bidding by Reinforcement Learning in Display Advertising by Han Cai et al. WSDM 2017.
Managing Risk of Bidding in Display Advertising by Haifeng Zhang et al. WSDM 2017.
Optimized Cost per Click in Taobao Display Advertising by Han Zhu et al. ArXiv 2017.
Combining Powers of Two Predictors in Optimizing Real-Time Bidding Strategy under Constrained Budget by Chi-Chun Lin et al. CIKM 2016.
Joint Optimization of Multiple Performance Metrics in Online Video Advertising by Sahin Cem Geyik et al. KDD 2016.
Optimal Real-Time Bidding for Display Advertising by Weinan Zhang. PhD Thesis 2016.
Bid-aware Gradient Descent for Unbiased Learning with Censored Data in Display Advertising by Weinan Zhang, Tianxiong Zhou, Jun Wang, Jian Xu. KDD 2016.
Lift-Based Bidding in Ad Selection by Jian Xu et al. AAAI 2016.
Feedback Control of Real-Time Display Advertising by Weinan Zhang et al. WSDM 2016.
Optimal Real-Time Bidding Strategies by Joaquin Fernandez-Tapia, Olivier Guéant, Jean-Michel Lasry. ArXiv 2015.
Programmatic Buying Bidding Strategies with Win Rate and Winning Price Estimation in Real Time Mobile Advertising by Xiang Li and Devin Guan. PAKDD 2014.
Statistical modeling of Vickrey auctions and applications to automated bidding strategies by Joaquin Fernandez-Tapia. Working paper.
~~Statistical Arbitrage Mining for Display Advertising by Weinan Zhang, Jun Wang. KDD 2015.非线性出价函数，优化目标：DSP净利润~~
Real-Time Bidding rules of thumb: analytically optimizing the programmatic buying of ad-inventory by Joaquin Fernandez-Tapia. SSRN 2015.
~~Optimal Real-Time Bidding for Display Advertising by Weinan Zhang, Shuai Yuan, Jun Wang. KDD 2014.非线性出价函数，优化目标：点击数~~
~~Bid Optimizing and Inventory Scoring in Targeted Online Advertising by Claudia Perlich et al. KDD 2012.M6d线性出价函数~~
Real-Time Bidding Algorithms for Performance-Based Display Ad Allocation by Ye Chen et al. KDD 2011

CTR/CVR Estimation

Deep & Cross Network for Ad Click Predictions by Ruoxi Wang et al. AdKDD & TargetAd 2017.
Deep Interest Network for Click-Through Rate Prediction by Guorui Zhou et al. ArXiv 2017.
DeepFM: A Factorization-Machine based Neural Network for CTR Prediction by Huifeng Guo et al. IJCAI 2017
Learning Piece-wise Linear Models from Large Scale Data for Ad Click Prediction by Kun Gai, Xiaoqiang Zhu, Han Li, et al. Arxiv 2017.
SEM: A Softmax-based Ensemble Model for CTR Estimation in Real-Time Bidding Advertising by Wen-Yuan Zhu et al. BigComp 2017.
Neural Feature Embedding for User Response Prediction in Real-Time Bidding (RTB) by Enno Shioji, Masayuki Arai. ArXiv 2017.
Field-aware Factorization Machines in a Real-world Online Advertising System by Yuchin Juan, Damien Lefortier, Olivier Chapelle. ArXiv 2017.
Product-based Neural Networks for User Response Prediction by Yanru Qu et al. ICDM 2016.
Sparse Factorization Machines for Click-through Rate Prediction by Zhen Pan et al. ICDM 2016.
Deep CTR Prediction in Display Advertising by Junxuan Chen et al. MM 2016.
Bid-aware Gradient Descent for Unbiased Learning with Censored Data in Display Advertising by Weinan Zhang, Tianxiong Zhou, Jun Wang, Jian Xu. KDD 2016.
Large Scale CVR Prediction through Dynamic Transfer Learning of Global and Local Features by Hongxia Yang et al. BIGMINE 2016.
Predicting ad click-through rates via feature-based fully coupled interaction tensor factorization by Lili Shan, Lei Lin, Chengjie Sun, Xiaolong Wang. Electronic Commerce Research and Applications 2016.
- Simple and Scalable Response Prediction for Display Advertising by Olivier ChapelleCriteo, Eren Manavoglu, Romer Rosales. ACM TIST 2014.
Cost-sensitive Learning for Utility Optimization in Online Advertising Auctions by Flavian Vasile, Damien Lefortier, Olivier Chapelle. Extension under-review of the paper presented at the Workshop on E-Commerce, NIPS 2015.
User Response Learning for Directly Optimizing Campaign Performance in Display Advertising by Kan Ren, Weinan Zhang, Yifei Rong, Haifeng Zhang, Yong Yu, Jun Wang. CIKM 2016.
A Convolutional Click Prediction Model by Qiang Liu, Feng Yu, Shu Wu, Liang Wang. CIKM 2015.
Factorization Machines with Follow-The-Regularized-Leader for CTR prediction in Display Advertising by Anh-Phuong Ta. BigData 2015.
Deep Learning over Multi-field Categorical Data: A Case Study on User Response Prediction by Weinan Zhang, Tianming Du, Jun Wang. ECIR 2016.
Offline Evaluation of Response Prediction in Online Advertising Auctions by Olivier Chapelle. WWW 2015.
Predicting Response in Mobile Advertising with Hierarchical Importance-Aware Factorization Machine by Richard J. Oentaryo et al. WSDM 2014.
Scalable Hierarchical Multitask Learning Algorithms for Conversion Optimization in Display Advertising by Amr Ahmed et al. WSDM 2014.
Estimating Conversion Rate in Display Advertising from Past Performance Data by Kuang-chih Lee et al. KDD 2012.
Scalable Hands-Free Transfer Learning for Online Advertising by Brian Dalessandro et al. KDD 2014.
Evaluating and Optimizing Online Advertising: Forget the click, but there are good proxies by Brian Dalessandro et al. SSRN 2012.
Modeling Delayed Feedback in Display Advertising by Olivier Chapelle. KDD 2014.
Ad Click Prediction: a View from the Trenches by H. Brendan McMahan. KDD 2013.
Practical Lessons from Predicting Clicks on Ads at Facebook by Xinran He et al. ADKDD 2014.

2017

Attribution Modeling Increases Efficiency of Bidding in Display Advertising by Eustache Diemert et al. ADKDD 2017.

归属模型提高展示广告竞价效率

数据集：Criteo、Facebook真实场景

offline数据结：http://research.criteo.com/criteo-attribution-modeling-bidding-dataset/

Abstra

ct以往算法将转化率预估和转化行为归因于哪次广告行为（展示或者点击）分开研究，是两个独立的问题，但是作者发现，归因模型有助于提高竞价的效果，因此将转化行为归因于哪次广告活动加入竞价策略中，提高竞价效率。

Introduction

衡量展示广告效果基于广告展示带来的价值，目前主要的两种方式：CPC，CPA（cost per action）,工业界中归因转化的标准是：归因于转化之前30天之内最后一次点击行为。

对于广告引擎和平台来说，目前最先进的出价策略是：Expected Value Bidder (EVB),也就是根据展示机会的预期价值来出价。预期价值是广告主payment*预测的转化率

在博弈论和经济学领域，转化归因已有深入研究，已经有一些机制被提出去较好地衡量广告主的payment，尤其是在多个广告点击或者转化行为同时发生或多个渠道同时存在的情况下。

考虑一个广告行为序列，合理的出价方式是后面的展示行为都应该比第一次展示的价值低，因为第二次点击行为能够导致转化的可能性不高。在之前的EVB策略中，是没有这方面考虑的，因为它既不能跟踪也不能预测归因转化率，所以为了改进实时竞价策略，我们建议利用一个归因模型来修改投标策略。

Model

把转化归因于最后一次点击行为适用于CPM结算方式，但不是很适合CPCorCPA方式。

从数据中学习归因关系，假设广告投放平台可以访问其广告主客户的转化归属标签，（该广告主通过其他广告投放平台的转化归主标签不可访问）

将归属关系表达成间隔时间的负指数函数，竞价策略为CPA*归属概率。

Data description

This dataset represents a sample of 30 days of Criteo live traffic data. Each line corresponds to one impression (a banner) that was displayed to a user. For each banner we have detailed information about the context, if it was clicked, if it led to a conversion and if it led to a conversion that was attributed to Criteo or not. Data has been sub-sampled and anonymized so as not to disclose proprietary elements.

Here is a detailed description of the fields (they are tab-separated in the file):

timestamp: timestamp of the impression (starting from 0 for the first impression). The dataset is sorted according to timestamp.
uid a unique user identifier
campaign a unique identifier for the campaign
conversion 1 if there was a conversion in the 30 days after the impression (independently of whether this impression was last click or not)
conversion_timestamp the timestamp of the conversion or -1 if no conversion was observed
conversion_id a unique identifier for each conversion (so that timelines can be reconstructed if needed). -1 if there was no conversion
attribution 1 if the conversion was attributed to Criteo, 0 otherwise
click 1 if the impression was clicked, 0 otherwise
click_pos the position of the click before a conversion (0 for first-click)
click_nb number of clicks. More than 1 if there was several clicks before a conversion
cost the price paid by Criteo for this display (disclaimer: not the real price, only a transformed version of it)
cpo the cost-per-order in case of attributed conversion (disclaimer: not the real price, only a transformed version of it)
time_since_last_click the time since the last click (in s) for the given impression
cat[1-9] contextual features associated to the display. Can be used to learn the click/conversion models. We do not disclose the meaning of these features but it is not relevant for this study. Each column is a categorical variable. In the experiments, they are mapped to a fixed dimensionality space using the Hashing Trick (see paper for reference).

Criteo Data

timestamp: 时间戳
uid 用户id
campaign 推广计划id
conversion 转化
conversion_timestamp 转化时间
conversion_id 转化id
attribution 转化归属关系
click 点击
click_pos 点击次序
click_nb 点击数量
cost 成本价格，Criteo赢得展示所支付的费用
cpo the cost-per-order in case of attributed conversion (disclaimer: not the real price, only a transformed version of it)
time_since_last_click 最后一次点击时间到转化的间隔时间
cat[1-9] 上下文特征，hash过。