cikm 2013 tutorial: real-time bidding: a new frontier of computational advertising research
DESCRIPTION
Computational Advertising has been an important topical area in information retrieval and knowledge management. This tutorial will be focused on real-time advertising, aka Real-Time Bidding (RTB), the fundamental shift in the field of computational advertising. It is strongly related to CIKM areas such as user log analysis and modelling, information retrieval, text mining, knowledge extraction and management, behaviour targeting, recommender systems, personalization, and data management platform. This tutorial aims to provide not only a comprehensive and systemic introduction to RTB and computational advertising in general, but also the emerging research challenges and research tools and datasets in order to facilitate the research. Compared to previous Computational Advertising tutorials in relevant top-tier conferences, this tutorial takes a fresh, neutral, and the latest look of the field and focuses on the fundamental changes brought by RTB. We will begin by giving a brief overview of the history of online advertising and present the current eco-system in which RTB plays an increasingly important part. Based on our field study and the DSP optimisation contest organised by iPinyou, we analyse optimization problems both from the demand side (advertisers) and the supply side (publishers), as well as the auction mechanism design challenges for Ad exchanges. We discuss how IR, DM and ML techniques have been applied to these problems. In addition, we discuss why game theory is important in this area and how it could be extended beyond the auction mechanism design. CIKM is an ideal venue for this tutorial because RTB is an area of multiple disciplines, including information retrieval, data mining, knowledge discovery and management, and game theory, most of which are traditionally the key themes of the conference. As an illustration of practical application in the real world, we shall cover algorithms in the iPinyou global DSP optimisation contest on a production platform; for the supply side, we also report experiments of inventory management, reserve price optimisation, etc. in production systems. We expect the audience, after attending the tutorial, to understand the real-time online advertising mechanisms and the state of the art techniques, as well as to grasp the research challenges in this field. Our motivation is to help the audience acquire domain knowledge and obtain relevant datasets, and to promote research activities in RTB and computational advertising in general.TRANSCRIPT
Real-Time BiddingA New Frontier of Computational Advertising Research
Jun Wang and Shuai Yuan, University College London
Xuehua Shen, iPinyou
Samuel Seljan, AppNexus
About us
• Dr Jun Wang and Shuai Yuan from University
College London
– Media future research group
– Computational advertising (big data analytics
and web economics)
• Dr Jun Wang is a Senior Lecturer (Associate
Professor) of Department of Computer Science,
UCL
• Shuai Yuan is Jun’s PhD student in CRS
(completing research status)
2
About us contd.
• Dr Xuehua Shen is CTO and co-founder of
iPinYou
– He received his PhD of Computer Science at
University of Illinois at Urbana-Champaign,
USA
iPinYou is the largest Demand Side Platform (DSP) and the leader of audience
targeting and real-time advertising in China. It makes intelligent decision for more
than 3 billion ads impressions each day. In the past two years, iPinYou is the pioneer
of programmatic buying of display media in China and organizes the annual RTB
Summit. It has more than 150 employees and is headquartered in Beijing and has
offices in Shanghai, Guangzhou, and Silicon Valley.
3
About us contd.
• Dr Samuel Seljan is a Quantitative Analyst at AppNexus
– Supply-side optimization
• to improve the allocation of impressions across RTB and non-RTB markets
• reserve price optimization
– He obtained a PhD in Political Science from the University of California, San
Diego
• AppNexus is one of the largest real-time advertising platforms (exchanges)
– Offers one of the most powerful, open and customizable advertising technology
platforms
– Serves Microsoft Advertising Exchange, Interactive Media (Deutsche Telekom),
and Collective Exchange
4
Outline• The background of RTB (25min)
– history, glossaries, fundamental challenges, players and their objectives
• An empirical study of RTB auctions (15min)
– periodic features, bids’ distribution, daily pacing, frequency & recency control
• Demand side optimisation (40min)
– bidding algorithms, conversion attribution
• The iPinyou global bidding algorithm competition (30min, break 30min)
– results, prizes, how to participate
• Supply side optimisation (40min)
– ad density, reserve prices, revenue channel selection, bid landscape
forecasting, pricing guaranteed delivery, data leakage
• Financial methods in computational advertising (15min)
– game theory basic, ad options
• Panel discussion (15min)5
Part 1: The background of RTB
• Egyptians used papyrus to make sales
messages and wall posters (4000 BCE)
• In the 18th century, ads started to appear
in weekly newspapers in England
• Thomas J. Barratt has been called "the
father of modern advertising"
1806 1890
1900
courtesy of Wikipedia 6
1952
Ads can be not annoying
7courtesy of lostateminor.com
Glossaries
• Real-Time Bidding is an important aspect of Programmatic buying, which is
getting more and more popular in Display (related) advertising. Another
major part of Online advertising is Sponsored search
• An Impression is an ad display opportunity which generates when a User
visits a webpage containing ad Placements
• The Publisher sends a bid request of this impression to an Ad network, or an
Ad exchange via his Supply side platform (SSP), then to Demand side
platforms (DSP) to reach Advertisers
• Usually, DSPs contact Data management platform (DMP) to check the
Segments of the current user, i.e., his intents or interests. Then a bid will be
computed for the Campaign
• The payment among these entities is usually in Cost per mille (CPM), but
sometimes could be Cost per click (CPC) or Cost per acquisition (CPA)
• If the advertiser wins the impression, his Creative will be displayed to the
user8
The fundamental challenges
• To find the best match between a given user in a given context and a
suitable advertisement?
• To achieve the best campaign performance (e.g., ROI) within the budget
constraint?
• To generate the most revenue given the traffic and demand?
• To maintain a healthy environment so that users get less annoyed (both
quality and quantity)?
Computational advertising, AZ Border, 2008Dynamics of bid optimization in online advertisement auctions, C Borges et al. 2007Dynamic revenue management for online display advertising, G Roels and K Fridgeirsdottir, 2009Advertising in a pervasive computing environment, A Ranganathan and RH Campbell, 2002 9
The simplified history of online (display) advertising
Real-time Bidding for Online Advertising: Measurement and Analysis, S Yuan et al., 2013 10
Direct sales
27th Oct 1994, AT & T on HotWired.com(78% CTR)
• Advertisers and publishers talk to (4A) agencies
• Still popular in today’s marketplace
courtesy of Ad Age 11
Trading in ad networks
courtesy of Admeld
Why?• After direct sales, some impressions
will remain unsold (remnants)• Small publishers cannot find buyers
directly
Ad networks are first-level aggregators of (long-tail) demand and supply.
12
Introducing the ad exchange
courtesy of www.liesdamnedlies.com
single ad network is easy a few ad networks are manageable
hundreds of ad networks are nightmare13
A video
courtesy of Internet advertising bureau, src: http://www.youtube.com/v/1C0n_9DOlwE 14
Introducing the ad exchange contd.
• Ad exchanges are
marketplaces
• Advertisers and
publishers have to rely
on tools to connect
• Real-Time Bidding
promotes user-oriented
bidding
Ad exchanges are second-level aggregators of demand and supply
15
The complex display ad eco-system
courtesy of LUMAscape 2011 16
A new picture in 2013
courtesy of Rare Crowd
AggregatorsDemand side Supply side
17
Boundaries are getting blurry
courtesy of Google
Google is introducing display ads to search result pages
18
Introducing the Demand Side Platform (DSP)
courtesy of LUMAscape 2011
• To connect to ad exchanges and SSPs
• To buy user-data from DMPs
• To provide campaign management
functions
• To bid by targeting rules and
optimisation algorithms
• To report and analyse the performance
A demand side platform (DSP), also called buy side optimizer and buy side platform is a technology platform that provides centralized and aggregated media buying from multiple sources including ad exchanges, ad networks and sell side platforms, often leveraging real time bidding capabilities of these sources.
IAB Wiki
19
DSP contd.
Bidding algorithm is the core of a DSP
20
Introducing the Supply Side Platform (SSP)
courtesy of LUMAscape 2011
• To upload advertisements and rich media
• To traffick ads according to differing business
rules
• To target ads to different users, or content
• To tune and optimise
• To report impressions, clicks, post-click &
post-impression activities, and interaction
metrics
A sell side platform (SSP), also called sell side optimizer, inventory aggregator, and yield optimizer is a technology platform that provides outsourced media selling and ad network management services for publishers.
IAB Wiki
21
SSP contd.
Yield optimisation is the core of a SSP
22
Introducing the Data Management Platform (DMP)
• To collect users’ online behaviour data across
websites
(Mainly via 3rd party cookies)
• To predict users’ segments (intents/interests)
bases on online behaviour data
• To answer the query of users’ segments
• To provide audience profiling and expansion
services
courtesy of LUMAscape 2011
A Data Management Platform (DMP) is a system that allows the collection of audience intelligence by advertisers and ad agencies, thereby allowing better ad targeting in subsequent campaigns.
IAB Wiki
23
DMP contd.
The user base and the learning engine are the cores of a DMP
24
Behind the banner
cmsummit.com/behindthebanner
Behind the banner
(A visualization of the adtech ecosystem)
Adobe, 2013
25
Part 2: An empirical study of RTB
• To understand the bidding behaviours in RTB auctions
• To present some research challenges
• To help to get familiar with RTB in the real-world
• The data is from production DSP & SSP based in UK
– 52m impressions, 72k clicks, and 37k conversions from Feb to May 2013
• Started from convs/clicks and back-traced to imps
– 12m auctions from 50 placements from Dec 2012 to May 2013
• 16 websites of different categories
26
Periodic patterns
The numbers of imp (left) and click (right) both show strong daily and weak weekly patterns,corresponding to the normal human activity
27
Periodic patterns contd.
Daily periodic patterns for conv (left) and cvr (right) show thatpeople are less likely to convert during late night
28
Frequency distribution
The frequency against CVR plot from two different campaignsCampaign 1 sets a frequency cap of 2-5 -> poor performanceCampaign 2 sets a frequency cap of 6-10 -> waste of budget
29
Recency distribution and conversion window
The recency factor affects the CVR (left)Campaign 1 sets a long recency cap -> waste of budgetCampaign 2 sets a short recency cap -> poor performance
The wide conversion window (right) challenges attribution models
30
Periodic patterns contd.
The winning bids peak at 8-10am due to intensive competition
31
Level of competition(number of bidders)
Change of winner
The more bidders, the higher chance of winner change, which makes it harder to detect a dynamic reserve price
32
Bids’ distribution
Accepted (p>0.05) Rejected
AD test per auction 0.343 0.657
AD test per placement 0.000 1.000
CQ test per auction 0.068 0.932
The commonly adopted assumption of Uniform distribution or Log-normal distribution
were mostly rejected
• Anderson-Darling test for Normality
• Chi-squared test for Uniformity
Finding the best fit of bids’ distribution is important:
• Optimal reserve price
• Bid landscape forecasting
• etc.
And what’s the granularity? (placement, geographical location, time & weekday, etc.)33
Budgeting and daily pacing
Real Time Bid Optimization with Smooth Budget Delivery in Online Advertising, KC Lee et al., 2013 34
Budgeting and daily pacing
35
A mixture of 1st and 2nd price auctions
• A high soft floor price can make it 1st price auction(In RTB, floor prices are not always
disclosed before auctions)
• In our dataset, 45% 1st price auctions consumed 55% budgets
• The complicated setting puts advertisers in an unfavourable position and could damage the ad eco-system
36
Overview: references
• The History of Advertising: How Consumers Won the War for Their Attention, HubSpot, 2013blog.hubspot.com/the-history-of-advertising-war-for-consumer-attention-slideshare
• How Cluttered Is the Advertising Landscape, Really? HubSpot, 2013blog.hubspot.com/how-cluttered-is-advertising-landscape-timeline
• Navigating Planet Ad Tech, MIT Technology Review, 2013www.technologyreview.com/view/518551/the-evolution-of-ad-tech/
• Internet Advertising: An Interplay among Advertisers, Online Publishers, Ad Exchanges and Web Users, S Yuan et al., 2013arxiv.org/abs/1206.1754
• Ad exchanges: research issues, S Muthukrishnan, 2009sites.google.com/site/algoresearch/start2.pdf
• Behind the banner (A visualization of the adtech ecosystem), Adobe, 2013cmsummit.com/behindthebanner/
37
Part 3: Demand side optimisation
• Bid optimisation
• Conversion attribution
38
Bid optimisation
• Input
– logs for auctions, impressions and
events
– targeting rules
– budgets and pacing preference
– internal/external user data
• The Decision Engine
– Gradient Boosting Regression
decision Tree, etc.
– fast & scalable
• Output
– to bid or not
– how much
39
Bid optimisation
• Baseline (constant or random, for exploration)
• Simple rule based (to bid high if the return is high)
– bid = base * pred_CTR / avg_CTR
– bid = conv_value * CVR * ROI
• Regression for estimation
– Generalised linear regression models (logistic, Bayesian probit, FTRL-Proximal,
etc.)
– Tree based models (random forest, gradient boosting regression tree, etc.)
– Neural networks and deep learning
40
Bid optimisation
• Looks good, but…
– metrics for evaluation?
– exploration vs.. exploitation (esp. for cold-start campaigns)
– risks (variance or confidence intervals from estimation)
– practical constraints (branding, overspending risks, inconsistent billing units, etc.)
• E-E problem
– Interactive collaborative filtering
– dimension deduction, correlation, etc.
• Risks
– Defining the Utility as the objective
– Portfolio theory
41
Metrics
• Top funnel metrics (to gain brand awareness)
– brand recall (awareness uplift)
– branded search
– direct website traffic
• Mid funnel metrics (to educate and engage the prospects)
– cost per new website visitor
– page view & form uplift
• Bottom funnel metrics (to generate value both online and offline)
– total conversion
– cost per conversion
– opportunity contribution (interested but not converted yet)
– revenue
courtesy of Adexchange 42
Transfer learning
• The Problem
– CTR is no good metrics but CVR is
too low
• Task
– To train on site visits
• Challenge
– Which site visits, and weight?
– Data availability
• Solution
– Similarity (contextual as a priori,
Bayesian)
Evaluating and Optimizing Online Advertising: Forget the Click, But There are Good Proxies, B Dalessandro, 2012 43
Conversion attribution problem
Dr Samuel Seljan from AppNexus
44
AppNexus
• AppNexus – Open and customizable advertising technology platform
• Process over 50 Billion ad requests per day
• Allow buyers to buy from over 90% of the web, including Facebook
• Clients are advertisers, ad agencies, content providers, and ad networks
• Major clients include:
• Microsoft
• Netflix
• eBay
• Zynga
• Interactive Media
• Orange - European telecom/media conglomerate
• WPP - world’s second largest ad agency
45
Why advertise?
“Done well, advertising sends a whisper to your
impulses – a primal wind at the back of your neck,
suggesting where to go and what to do.”
-Don Draper
46
Fundamental question of advertising
• But, does advertising work?
• Which ads work with what audience?
• Do the benefits of advertising outweigh it’s costs?
– Incremental revenue > Marginal cost?
47
Fundamental question of advertising
• The promise of digital advertising: precise measurement of users’ responses
to ads
• 1990s: Click tracking
• Compare CTRs across many, many dimensions:
• Campaign, image, time of day, region, location on page, gender, etc.
• Limitations?
48
Fundamental question of advertising
• The promise of digital advertising: precise measurement of users’ responses
to ads
• 1990s: Click tracking
• Compare CTRs across many, many dimensions:
• Campaign, image, time of day, region, location on page, etc
• Limitations?
• Late 2000s: Conversion tracking
• Conversion: when a user sees and ad and then takes an action, e.g. buys a pair or shoes
• Cost Per Action (CPA) payment: advertisers only pay when a conversion occurs
• Traders (agents of buyers) or sellers take on all the risk
• Does this answer the fundamental question of advertising?
49
Problems with CPA advertising
1. Users often see many ads for the same brand on many sites
2. Only includes online actions
3. Causal inference – what does the association between seeing an ad and
converting mean?
4. Ration of conversions per ads is often very small
– Between 1-5 conversions per 100,000 impressions is common
– Thus, takes many ads to learn “true” conversion rate.
• For a one month campaign with a $5,000 budget, possible to learn conversion rate with
5% error on roughly 10 different web sites!
– Difference between 1 and 5 conversions per 100,000 impression is difference
between a profitable and unprofitable campaign
50
Problems with CPA advertising
1. Users often see many ads for the same brand on many sites
• Which ad “is responsible” for the conversion?
• Industry standard is “last touch” attribution
– Previous graph is misleading!
– Doesn’t show how many times a user saw ad
• Connection to frequency optimization: creates a bias towards higher frequency ads
– Last touch is the most recent ad, but we don’t know for sure that user even saw the
most recent ad – we just have a record of it
– If effect of seeing ad is cumulative, this under weights importance of first view
51
Problems with CPA advertising
1. Users often see many ads for same brand on many sides
– Which ad and site gets credit for the conversion?
2. No tracking of offline purchases
– Technological and privacy challenges
• 3. The fundamental problem of causal inference
(Ad tech version)
Incremental revenue from ad j for user i =
Lifetime revenuei | i sees j– Lifetime revenuei | i does not see j
– But, each user either sees or does not see an ad so this cannot be calculated, even with
limitless data!
– Moreover, advertisers target users that are more likely to buy! (retargeting), thus lifetime
revenuei | i sees j is very likely to overestimate incremental revenue.
• Most CPA optimization creates selection bias
Problems with CPA advertising
• Alternative framing: most CPA buying algorithms are predicting who is most
likely to buy and then focusing (targeting) buying on these users
• Thus, advertisers may see a lot of conversions associated with ads on some sites
on some users, but do not know how much revenue they would have had without
those ads
• Example: amazon targets ads at people who have recently searched for an item
on their website.
• These people are more likely to buy on Amazon than those who have not recently
searched
• But, they are also more likely to buy without seeing an ad
Problems with CPA advertising
• Problem is more important for some brands than others
• For internet advertising, the potential scale of the problem can be considered by
thinking about:
Lifetime revenuei | i does not see digital ad
• For what types of campaigns is this likely to be a big problem? A small problem?
• Principal agent problem: the people that need to understand this are the
brands themselves
• Agents buying for brands, do not have a short term incentive to solve problem
– they get paid per conversion!
• Selling more rigorous CPA optimization to brands is challenging
54
Solutions to problems with CPA advertising
Problem 1: Many ads per conversion
• AppNexus Solution 1: custom conversion attribution
• We track conversions, but allow clients to divide conversion among ads
using their own “secret sauce”
• AppNexus Solution 2: Conversion Funnel
• Use events higher up the funnel to predict final sale
56
Problem 2: Offline tracking
• AppNexus Solution 1: allow for integrations with offline data providers and
the insertion of external data into optimization
• Other solutions?
• A privacy quid pro quo? Data for savings?
57
Problem 3: Fundament Problem of Causal Inf.
• AppNexus Solution: Random assignment
• Randomly assign users to group A (sees ad) or B (doesn’t see ad)
• Estimate incremental revenue from ad as:
conversion value * p(Conversion |A ) – p(Conversion | B)
– limitations: scalability, does not include off-line revenue or model social returns
• This is a problem for both group A and group B for larger brands
58
Problem 3: Fundament Problem of Causal Inf.
• Gold standard:
• Fowler et al. (2012): Facebook “I voted” and turnout
– RA to three groups:
» 1. Top panel
» 2. Second panel
» 3. No message
• Used federal data to compare turnout rates in groups & friends of those in each gorup
– .4% higher turnout of second gorup
• Combines solutions 2 and 3 (randomization addresses problems 2 and 3!)
59
Problem 3: Fundament Problem of Causal Inf.
• Remaining questions for RA: how to combine RA with other elements of a
CPA optimization algorithm…
– Over how many groups should one randomize?
» E.g. it could solve the frequency – conversion attribution problem, but that’s
a lot of groups
– What percent of impressions should be in test and control group?
60
Demand side optimisation: references
• Optimal bidding on keyword auctions, B Kitts and B Leblanc, 2004
• Stochastic gradient boosted distributed decision trees, J Ye et al., 2009
• Web-scale bayesian click-through rate prediction for sponsored search advertising in Microsoft's Bing search
engine, T Graepel et al., 2010
• Web-search ranking with initialized gradient boosted regression trees, A Mohan et al., 2011
• A Gentle introduction to random forests, ensembles, and performance metrics in a commercial system, D
Benyamin, 2012
citizennet.com/blog/2012/11/10/random-forests-ensembles-and-performance-metrics/
• A 61-million-person experiment in social influence and political mobilization, RM Bond et al., 2012
• Deep metworks for predicting ad click through rates, G Corrado, 2012
• Click modeling for display advertising, O Chapelle, 2012
• Causal reasoning and learning systems, L Bottou and E Portugaly, 2012
• Ad click prediction: a view from the trenches, HB McMahan et al., 2013
• Deep learning, yesterday, today, and tomorrow, K Yu et al., 2013
• Deep learning of representations: looking forward, Y Bengio, 2013
61
iPinyou global bidding algorithm competition
Dr Xuehua Shen from iPinyou
62
Best Algorithm
Maximize #clicks + N * #conversions
Subject to the fixed budget
3 Milestones +
1,000,000
Grand Prize
iPinYou Contest vs. Netflix Prize
Open Question
Offline & Online Evaluation
Production setting
Meaningful metrics
Dynamic data set
Short time span
7.5Gbid: 13.6Mimp: 9.2Mclk: 7.5Kconv: 72
Dropbox: https://www.dropbox.com/sh/xolf5thu8jsb
mfu/kBrAsSxtAN
百度网盘: http://pan.baidu.com/share/link?shareid=
374646&uk=3037373637
April 20 May 3 May 16
2000+
submissions
05.16 ~ 05.22 Warmup
05.23 ~ 05.25 Bidding
ml_rush, the9thbit, newline
Not just CTR
June 1 ~ August 31 Offline
Sept 1 ~ Sept 30 Online
Sept 6 ~ Sept 12 Warmup
Sept 13 ~ Sept 15 Bidding
Season 2 vs. Season 1
Support Python, R, Java
Competition platform
Top 5 go to Online stage
More bonus
Team membership (6 to 10)
2x DataUser Profile
July 12July 5 Aug 31 Aug 31’
4000+
submissions
Bidathon 2
RankTeam name
#Bid #Imp #Clk#Con
Spending
Win Rate
CTRFinal Score
1 o_o 50,874,976 4,658,979 4,265 4 1497.55 9.2% 0.092% 4289
2梦想照进现实
79,426,606 7,196,004 4,061 1 1481.59 9.1% 0.056% 4067
3 UCL-CA 3,597,284 2,059,509 1,989 1 1500.01 57.3% 0.097% 1995
4 Again 21,099,070 3,503,178 1,397 2 1499.99 16.6% 0.040% 1409
5 deep_ml 4,682,521 1,886,929 1,142 1 1500.08 40.3% 0.061% 1148
Bid More, Bid
Less
Season 3Oct 1 ~ Nov 15 Offline
Nov 15 ~ Dec 15 Online
Season 3 vs. Season 1/2
Even bigger data set
Focus on Online Stage
Mobile Campaign
Any question sent to
全球RTB算法大赛
320076711
iPinYou Global RTB Bidding
Algorithm Competition
Part 4: Supply side optimisation
• Typical revenue models
• Ad density optimal control
• Reserve price optimisation
• Ad channel selection
• Connecting the supply side markets
• Data leakage protection and pricing
103
Typical revenue models for the supply side
• Subscription access to content (FT.com)
• Pay Per View access to document (Downloading a paper outside the campus)
• CPM display advertising on site
• CPC advertising on site (Google AdSense)
• Sponsorship of site sections or content types (typically fixed fee for a period)
• Affiliate revenue (Compare shopping, CPA/CPC)
• Subscriber data access for marketing (VISA & MasterCard)
• User contributed data for marketing (Surveys)
Publishers will seek to use the best combination of these techniques
104
Ad density
The task:
• To find the optimal advertising
density (number of ad placements)
for a given website
The challenges:
• Users’ preference model
• Expected CPM
• Competition
The assumption:
• Using real-time bidding only
105
No ads
courtesy of www.gov.uk
Some websites do not rely on
ads to compensate the
maintenance cost
• Government
• Education
• Most of .org
106
All ads
• Created by Alex Tew in
2005
• Selling 100k 100-pixels
at $100 each
• Sold out in 4 months
• Almost 0% CTR
courtesy of www.milliondollarhomepage.com 107
Ad density: example
Apr 2010Sep 2013
courtesy of archive.web.org 108
Ad density: example
courtesy of Quantcast
Question: is it reasonable to put on so many more ads?
109
Ad density: an optimal control problem
• Assumptions
– Ad density and impressions determine revenue
– Ad density determines impressions
Ad density
Impressions (page views)CPM
Maintenance cost factor
Content attraction Ad repellence Natural growth
Management and valuation of advertisement-supported web sites, Dewan et al. 2005
Question: what’s the optimal densities of multiple publishers under competition?
110
Reference: ad density, layout & pricing
• Management and valuation of advertisement-supported web sites, RM Dewan, 2003
• Optimal pricing and advertising policies for web services, S Kumar et al., 2004
• Is revamping your web site worthwhile? EY Huang, 2005
• An economic analysis of ad-supported software, BJ Jiang, 2007
• Dynamic pricing and advertising for web content providers, S Kumar and SP Sethi, 2009
• Pricing display ads and contextual ads: Competition, acquisition, and investment, YM Li and JH Jhang-Li, 2009
• Dynamic ad layout revenue optimization for display advertising, H Cheng et al., 2012
• Automatic ad format selection via contextual bandits, L Tang et al., 2013
111
Reserve price optimisation
The task:
• To find the optimal reserve
prices
The challenge:
• Practical constraints v.s
common assumptions (bids’
distribution, bidding private
values, etc.)
The assumptions:
• 2nd price auction
• With only hard floor price
Even in the 2nd price auction, the winner does not always pay the 2nd highest bid (or minimal + $0.01)
112
Reserve price: flowchart
Suppose it is 2nd price auction
• Normal case: b2 > a• Preferable case: b1 > a > b2• Undesirable case: a > b1
113
• 2 bidders, Uniform[0, 1]
• Without a reserve price
• With the optimal auction theory
Reserve price: example
Reserve prices in internet advertising auctions: A field experiment, Ostrovsky and Schwarz, 2011 114
Reserve price: the optimal auction theory
• In the 2nd price auctions, advertisers bid their private values [𝑏1, … , 𝑏𝐾]
• Values are independently distributed and drawn from certain distributions
– Uniform
– Log-normal
• The publisher also has a private value 𝑉𝑝
• The optimal reserve price is given by
Questions:
• Are advertisers bidding their private values?
• Does Uniform/Log-normal fit well?
Optimal Reservation Prices in Auctions, Levin and Smith, 1996
𝐹 𝒃 = 𝐹1 𝑏1 ×⋯× 𝐹𝐾(𝑏𝐾)
𝛼 −1 − 𝐹 𝒃
𝐹′ 𝒃− 𝑉𝑝 = 0
115
Reserve price: field experiment results
Reserve prices in internet advertising auctions: A field experiment, Ostrovsky and Schwarz, 2011 116
Reserve price: field experiment results
Reserve prices in internet advertising auctions: A field experiment, Ostrovsky and Schwarz, 2011 117
Reserve price: detection and reaction
• A dynamic and repeated game between the winner (w) and the publisher (p)
• Extension form representation
– Information nodes:
• 𝐼1: the winning bid 𝑏1 is higher
• 𝐼2: the reserve price 𝛼 is higher
– Actions:
• 𝑎𝑤1: to increase 𝑏1 so that 𝑏1 ≥ 𝛼
• 𝑎𝑤2: to increase 𝑏1 so that 𝑏1 < 𝛼
• 𝑎𝑤3: to decrease 𝑏1 so that 𝑏1 ≥ 𝛼
• 𝑎𝑤4: to decrease 𝑏1 so that 𝑏1 < 𝛼
• 𝑎𝑝1: to increase 𝛼 so that 𝛼 ≥ 𝑏1
• 𝑎𝑝2: to increase 𝛼 so that 𝛼 < 𝑏1
• 𝑎𝑝3: to decrease 𝛼 so that 𝛼 ≥ 𝑏1
• 𝑎𝑝4: to decrease 𝛼 so that 𝛼 < 𝑏1
118
Reserve price: detection and reaction
The dynamic and static game tree for the auction with reserve price
payoff of the winner and the publisher
119
Reserve price: detection and reaction
• Consider playing the game repeatedly
– If the reserve price was higher, should the publisher lower it?
– The optimal auction theory
• Advertisers want to learn the publisher’s private value distribution, too
– What is the absolute minimal price that can be accepted?
• Questions
– What are the best response functions for both players?
– What are the dominant strategies and equilibrium?
120
Reserve price: references
• On optimal reservation prices in auctions, Engelbrecht-Wiggans, 1987
• Optimal reservation prices in auctions, Levin and Smith, 1996
• Auction theory: a guide to the literature, Klemperer, 1999
• Reserve prices in internet advertising auctions: a field experiment, Ostrovsky and Schwarz, 2009
• Auction theory 2nd edition, Krishna, 2009
• Optimal reserve price for the generalized second-price auction in sponsored search advertising, Xiao et al.,
2009
• Optimal auction design and equilibrium selection in sponsored search auctions, Edelman and Schwarz, 2010
• Optimal auction design in two-sided markets, R Gomes, 2011
121
• The task:
– There are multiple ad
channels giving
different payoffs over
time. Which one to
use?
• The challenge:
– Too many possible
candidates
– The payoffs change
over time
Ad channel selection
Sequential Selection of Correlated Ads by POMDPs, Yuan and Wang, 2012 122
Ad channel selection contd.
• A sequential selection problem
• Value iteration exact solution (high computational complexity)
• Multi-armed bandit approximation
Sequential Selection of Correlated Ads by POMDPs, Yuan and Wang, 2012 123
Sequential selection: references
• Dynamic programming, RE Bellman, 1957
• Multi-armed bandits and the Gittins index, P Whittle, 1980
• A survey of POMDP applications, AR Cassandra, 1998
• A survey of POMDP solution techniques, KP Murphy, 2000
• Finite-time analysis of the multi-armed bandit problem, P Auer, P. et al.,2002
• Multi-armed bandit algorithms and empirical evaluation, J Vermorel and M Mohri, 2005
124
Impression allocation between GD and NGD
• Sometimes, ad channels could have different attributes:
– GD: Guaranteed Delivery (contracts)
– NGD: Non-Guaranteed Delivery (auctions)
• Given a time window, the publisher decides
– To accept or reject a contract proposal
– To allocate impressions among multiple contracts and auctions
(NGD can be modelled as an already accepted contract with ∞ required
impressions and 0 under-delivery penalty)
Dynamic Revenue Management for Online Display Advertising, Roels and Fridgeirsdottir, 2008 125
Connecting GD and NGD
courtesy of Internet advertising bureau
Four (4) major types of inventories in Programmatic selling
Now they are priced differently and separately
Separated markets implies inefficiency and arbitrage
126
Automated (programmatic) guaranteed delivery
• AOL Upfront
– Will take effect January 1, 2014
– Two brands and five agencies have committed, around $10m for each agency
(undisclosed)
– AOL’s ad placements, e.g. The Huffington Post, TechCrunch and StyleList
– A private marketplace in the beginning
• Question
– Selling mechanisms (auction, queue, etc.)
– Inventory allocation and reserve prices
courtesy of AOL 127
Optimal pricing of a guaranteed delivery contract
• Suppose the publisher wants to sell 𝑆 impressions from time step 𝑇 + 1
– In advance: guaranteed delivery contracts
– On spot: RTB auctions (non-guaranteed, 2nd price auction)
• Consider the total demand as 𝑄
• The private value distribution of an advertiser is 𝐹 ⋅
• The utility of an advertiser is 𝑈𝑎 ⋅ = 𝐹 ⋅ + 𝑔 𝑡
– Willing to pay higher if could buy in advance
• The utility of the publisher is 𝑈𝑝 ⋅ = 𝑅 ⋅ + ℎ(𝑡)
– Willing to charge lower if could sell in advance
What is the optimal price at 0 < 𝑡 < 𝑇 + 1 ?
128
Guaranteed delivery: references
• Optimal dynamic auctions for display advertising, YJ Chen, 2009
• Pricing guaranteed contracts in online display advertising, V Bharadwaj, 2010
• Risk-aware revenue maximization in display advertising, A Radovanovic and WD Heavlin, 2012
• Optimal allocation for display advertising, H Rui et al., 2012
• A unified optimization framework for auction and guaranteed delivery in online advertising, K Salomatin,
2012
• Maximally representative allocations for guaranteed delivery advertising campaigns, RP McAfee, 2013
129
Bid landscape forecasting
courtesy of Google AdWords
The tool usually exists as a service provided to advertisers
130
Bid landscape forecasting
Bid landscape forecasting in online ad exchange marketplace, Y Cui et al,. 2011
• The task:
Given a campaign (a set of
targeting rules), what is the
bid-impression distribution
for a given venue (domain,
placement, etc.)?
• The challenge:
Forecasting for new &
changed campaigns
Forecasting the win rate for
unseen bids
131
Bid landscape forecasting
Ad impression forecasting for sponsored search, A Nath et al., 2013 132
A Generative Model based ad Impression Forecasting method
Bid landscape forecasting: references
• Bid landscape forecasting in online ad exchange marketplace, Y Cui et al., 2011
• Handling forecast errors while bidding for display advertising, KJ Lang et al., 2012
• Ad impression forecasting for sponsored search, A Nath et al., 2013
• Forecasting user visits for online display advertising, S Cetintas et al., 2013
• Predicting advertiser bidding behaviors in sponsored search by rationality modeling, H Xu et al., 2013
• Optimizing volume and frequency forecasts for an online video advertiser, J Talbot et al., 2013
133
Data leakage protection
• Every player in the ad eco-system realises the value of audience data
– who owns it?
– what’s its value?
• The data leakage problem
– buyers collect user data from premium website
– then retarget these users on cheap inventories
• The task
– To learn who is collecting user data on the webpage (piggybacks)
– To stop the unauthorized collection and ask for payment
• The challenge
– The optimal price: limiting the user data access will hurt the CPM
134
Part 5: Financial methods in Computational Advertising
• Game theory basics
• Ad options
135
Examples
A futures market in computer time, IE Sutherland, 1968
Financial methods and game theory have a long history in CS research
Auction for sharing compute resources
136
Examples contd.
• Amazon EC2 is a web service that provides resizable compute capacity in the cloud• In late 2009, Amazon announce its spot instances pricing system
Deconstructing Amazon EC2 Spot Instance Pricing, OA Ben-Yehuda, 2011
Auction and futures in cloud computing
137
Examples contd.
courtesy of Wall Street Journal
In the age of the Internet, fixed prices are a thing of the pastProfessor Oren Etzioni
138
Game theory basics
• A game is “a competitive activity … in which players contend with each other
according to a set of rules”
• A strategic game (with ordinal preferences) consists of
– a set of players
– for each player, a set of actions
– for each player, preferences over the set of action profiles
• The best response
• The Nash equilibrium of static games
An introduction to game theory, MJ Osborne, 2003 139
Example: the prisoner's dilemma
The mysterious benedict society and the prisoner’s dilemma, TL Stewart, 2009courtesy of Encyclopaedia Britannica, 2006
• A typical example of non-
cooperative game
• The strictly dominant strategy
for both players is to confess,
which also forms the Nash
equilibrium of the game
• When played iteratively, the
winning deterministic strategy is
tit-for-tat
(first cooperate, then subsequently
replicate an opponent's previous
action)
140
Example: The game of chicken
• While each player prefers not to
yield to the other, the worst possible
outcome occurs when both players
do not yield.
– The nuclear crisis
– The promise on advertising effects
– The bidding on similar audiences
simultaneously
Dare Chicken
Dare 0, 0 7, 2
Chicken 2, 7 6, 6
courtesy of Rebel Without a Cause 141
Example: The Cournot competition• Assumptions
– There is more than one firm and all firms produce a
homogeneous product, i.e. there is no product
differentiation;
– Firms do not cooperate, i.e. there is no collusion;
– Firms have market power, i.e. each firm's output decision
affects the good's price;
– The number of firms is fixed;
– Firms compete in quantities, and choose quantities
simultaneously;
– The firms are economically rational and act strategically,
usually seeking to maximize profit given their competitors'
decisions.
courtesy of palimpsestes.fr 142
Example: The Cournot competition
• 𝑝1, 𝑝2: prices
• 𝑞1, 𝑞2: quantities
• Firm 1’s profit:
Π1 = 𝑞1 𝑃 𝑞1 + 𝑞2 − 𝑐
• Firm 1’s best response function:
𝑅(𝑞2) =𝜕Π1𝜕𝑞2
• To obtain the equilibrium:
𝑅 𝑞2 = 𝑅 𝑞1
(the intersection)
courtesy of Wikipedia 143
Example: The Bertrand competition
• Assumptions
– Firms compete by setting prices simultaneously and
consumers want to buy everything from a firm with a
lower price
– If two firms charge the same price, consumers
demand is split evenly between them
courtesy of Wikipedia 144
Example: The Bertrand competition
both firms are pricing at marginal cost
courtesy of Wikipedia
• If both firms set equal prices
above marginal cost, firms
would get half the market at a
higher than MC price
• By lowering prices just slightly,
a firm could gain the whole
market
• Both firms are tempted to
lower prices as much as they
can
• It would be irrational to price
below marginal cost, because
the firm would make a loss 145
Application of the duopoly competition
• If capacity and output can be easily changed -> Bertrand model
• if output and capacity are difficult to adjust -> Cournot model
• Analogy
– Two publishers serving similar content
– Premium and long-tail publishers seeing similar users
– etc.
146
References: financial methods and game theory basics
• A policy framework for trading configurable goods and services in open electronic markets, S Lamparter, 2006
• Planning and pricing of service mashups, B Blau et al., 2008
• Web service derivatives, T Meinl and B Blau, 2009
• How to coordinate value generation in service networks, B Blau et al., 2009
• Enabling cloud service reservation with derivatives and yield management, T Meinl et al., 2010
• Web services advanced reservation contracts, C Weinhardt et al., 2011
• Finite automata play the repeated prisioners dilemma, A Rubinstein, 1986
• A course in game theory, MJ Osborne and A Rubinstein, 1994
• An introduction to game theory, MJ Osborne, 2004
• Sponsored search auctions: an overview of research with emphasis on game theoretic aspects, P Mailléet al.,
2012
• Repeated keyword auctions played by finite automata, W Ding et al, 2013
147
Ad options
• The task:
– To sell impressions in advance
(a natural extension to the programmatic
guarantee)
– Both party can choose to exercise or not
• The (pricing) challenge:
– Impression prices are volatile
– Non-storability: impressions cannot be
bought and kept
– Not just about the price movements: the
uncertainty of traffic volume, CTR and etc.
courtesy of Webscope from Yahoo! Labs 148
Ad options contd.
Advertisers Publishers
secure impressions delivery
reduce uncertainty in auctions
cap cost
sell the inventory in advance
have a more stable and predictable
revenue over a long-term period
increase advertisers’ loyalty
Benefits
149
Submits a request of guaranteed ad delivery for the keywords ‘MSc Web Science’, ‘MSc Big Data Analytics’ and ‘Data Mining’ for the future 3 month term [0, T], where T = 0.25.
Ad options contd.
Sells a list of ad keywords via a multi-keyword multi-click option
t = T
Timeline
search engineonline advertiser
t = 0Pays £5 upfront option price to obtain the option.
Multi-Keyword Multi-Click Advertisement Option Contract for Sponsored Search, B Chen et al., 2013
multi-keyword multi-click option (3 month term)
upfront fee
(m = 100)keywords list fixed CPCs
£5
‘MSc Web Science’ £1.80
‘MSc Big Data Analytics’
£6.25
‘Data Mining’ £8.67
150
Exercising the option
t = T
Timeline
Exercises 100 clicks of ‘MSc Web Science’ via option.
t = 0
Pays £1.80 to the search engine for each click until the requested 100 clicks are fully clicked by Internet users.
t = t1
Reserves an ad slot of the keyword ‘MSc Web Science’ for the advertiser for 100 clicks until all the 100 clicks are fully clicked by Internet users..
t = t1c
Multi-Keyword Multi-Click Advertisement Option Contract for Sponsored Search, B Chen et al., 2013 151
search engineonline advertiser
Not exercising the option
t = T
Timeline
If the advertiser thinks the fixed CPC £8.67 of the keyword ‘Data Mining’ is expensive, he/she can attend keyword auctions to bid for the keyword as other bidders, say £8.
t = 0
Pays the GSP-based CPC for each click if winning the bid.
t = …
Selects the winning bidder for the keyword ‘Data Mining’ according to the GSP-based auction model.
Multi-Keyword Multi-Click Advertisement Option Contract for Sponsored Search, B Chen et al., 2013 152
search engineonline advertiser
Ad options pricing
• Building blocks
– No-arbitrage [F Black and M Scholes1973; H Varian1994]
– Stochastic underlying keyword CPC [P Samuelson1965]
– Terminal value formulation
• Formula
– n=1, Black-Scholes-Merton European call
– n=2, Peter Zhang dual strike European call
– n>=3, Monte Carlo method
Multi-Keyword Multi-Click Advertisement Option Contract for Sponsored Search, B Chen et al., 2013 153
Ad options: references
• Option pricing: a simplified approach, J Cox et al., 1979
• Online ad slotting with cancellations, F Constantin, 2008
• A truthful mechanism for offline ad slot scheduling, J Feldman et al., 2008
• Selling ad campaigns: online algorithms with cancellations, M Babaioff et al., 2009
• Options, futures and other derivative securities (7th edition), J Hull, 2009
• Online advertisement service pricing and an option contract, Y Moon and CY Kwon, 2010
• Selling futures online advertising slots via option contracts, J Wang and B Chen, 2012
• Multi-keyword multi-click advertisement option contract for sponsored search, B Chen et al., 2013
154
Additional references
• Internet advertising and the generalized second price auction: selling billions of dollars worth of keywords, B
Edelman, 2005
• Price cycles in online advertising auctions, X Zhang and J Feng, 2005
• Budget optimization in search-based advertising auctions, J Feldman et al., 2007
• The economics of the online advertising industry, DS Evans, 2008
• Expressive banner ad auctions and model-based online optimization for clearing, C Boutilier et al., 2008
• Computational advertising, AZ Broder, 2008
• Algorithmic methods for sponsored search advertising, J Feldman and S Muthukrishnan, 2008
• Internet ad auctions: Insights and directions, S Muthukrishnan, 2008
• The online advertising industry: economics, evolution, and privacy, , DS Evans, 2009
• Ad exchanges: Research issues, S Muthukrishnan, 2009
• Adaptive bidding for display advertising, A Ghosh et al., 2009
• The arrival of real-time bidding, Google, 2011
• Algorithms and strategies for web advertising, P Papadimitriou, 2011
• OpenRTB API specification, IAB, 2012
155
Additional references
• Targeted, not tracked: client-side profiles and privacy-friendly behavioral advertising, M Bilenko and M
Richardson, 2012
• Computational advertising in social networks, A Bhasin, 2012
• Size, labels, and privacy in targeted display advertising, C Perlich, 2012
• Estimating conversion rate in display advertising from past erformance data, K Lee et al,. 2012
• Handling forecast errors while bidding for display advertising, KJ Lang et al., 2012
• Marketing campaign evaluation in targeted display advertising, J Barajas et al., 2012
• Ad exchange-proposal for a new trading agent competition game, M Schain and Y Mansour, 2012
• Auctions for online display advertising exchanges: approximations and design, S Balseiro et al., 2012
• Real-time bidding for online advertising: measurement and analysis, S Yuan et al., 2013
• Impression fraud in on-line advertising via pay-per-view networks, K Springborn and P Barford, 2013
• An overview of computational challenges in online advertising, RE Chatwin, 2013
• Competition and yield optimization in ad exchanges, SR Balseiro, 2013
• Internet advertising revenue report, IAB and PwC
156