SlideShare a Scribd company logo
Promoting Positive Post-Click
Experience for In-Stream Yahoo
Gemini Users
Janette Lehmann
Freie Universität Berlin
Berlin, Germany
Mounia Lalmas
Yahoo Labs
London, UK
Guy Shaked
Yahoo
Tel Aviv, Israel
Gabriele Tolomei
Yahoo Labs
London, UK
Fabrizio Silvestri
Yahoo Labs
London, UK
The Online Advertising Arena
PUBLISHER
Maximise revenue from
online advertising
The Online Advertising Arena
ADVERTISER
Maximise ROI from his ad
campaigns
PUBLISHER
Maximise revenue from
online advertising
The Online Advertising Arena
ADVERTISER
Maximise ROI from his ad
campaigns
PUBLISHER
Maximise revenue from
online advertising
USER
Show me some content, not
these #*!%ing ads
(… unless they are “interesting”)
The Online Advertising Arena
We want of course to maximise our revenue and
guaranteeing ROI to our advertisers without negatively
impacting on user experience
Native Ads on Yahoo Stream
Gemini is the Yahoo’s marketplace for native ads (mobile + desktop)
and sponsored ads (mobile)
…
Native Ads on Yahoo Stream
Gemini is the Yahoo’s marketplace for native ads (mobile + desktop)
and sponsored ads (mobile)
Based on eCPM (expected Cost Per Mille Impressions)
CPC
(Cost Per Click)
CPC is the amount of money (bid) the
advertiser is willing to pay for 1 click on its ad
eCPM is computed from the underlying pricing model
rank(ad) = eCPM(ad)
Ad Ranking
rank(ad) = CPC(ad) * eCTR(ad) [* 1000]
P(click|ad)
Beyond CTR Optimisation
Ranking via CTR optimisation is an attempt to
maximise revenue on the short-term
Trade-off between provider’s and advertiser’s need
What about the user?
Better to include a “quality” factor so to optimise for
ads having “high quality”
rank(ad) = CPC(ad) * eCTR(ad)
P(click|ad) P(hq|click)
* eHQ(ad)
How Do We Estimate Quality?
“High Quality” is a shorthand for “High Quality Post-
Click User Experience”
Intuitively a high quality ad is one whose
conversion rate is high
Estimating the probability of conversion is hard!
- Conversion data is unavailable for a large set of advertisers
- Conversion data is not missing at random
Use “proxy” metrics of post-click quality like dwell
time and bounce rate on the ad landing page
Post-Click Quality Metrics
tad-click tback-to-stream
dwell time = tback-to-stream – tad-click
bounce rate = P(dwell time < th)
th =
5 secs. (mobile)
12 secs. (desktop)
Is Dwell Time A Good Proxy?
mobile
200K ad clicks
Ø  It needs less time
to get the same
probability of a
second click
desktop (Yahoo toolbar)
30K ad clicks
Ø  23.3% of users visit other
websites than the ad landing
page before returning to Yahoo
Ø  this goes down to 7.4% for dwell
time up to 3 mins.
Probability of a second click
increases with dwell time
Optimise for High Quality Ads
Estimating P(hq|click) turns into estimating:
P(dwell time > tδ)
P(bounce rate < τβ)
Can we build a model which is able to predict if an
ad is high quality according to the definitions above?
è  high quality = high dwell time
è high quality = low bounce rate
or
rank(ad) = CPC(ad) * eCTR(ad)
P(click|ad) P(hq|click)
* eHQ(ad)
Binary Classification Problem
AD
1
2
…
i
…
m
ad landing page
Xi
class label
Yi
+1 if hq(i)
-1 o.w.
dwell time > tδ
or
bounce rate < τβ
f(Xi)
Feature-based representation
Binary Classification Problem
find the estimate which
best approximates
P(Yi | f(Xi))
Features
How do we represent an ad landing page Xi?
CONTENT
clickToCall
numClickable
tokenCount
numImages
…
isMobileOptimised
SIMILARITY
cosineCreativeLandingPageNouns
cosineCreativeLandingPageWiki
HISTORICAL
numImpressions
numClicks
bouncerate
avgDwellTime
avgDwellTimeNonShort
ctr
cpc
Offline Experiments: Setup(1)
*Class probability obtained using Platt scaling
3 Learning Methods
- LogReg: L1-norm; C=100; stopping criterion = 10-2
- SVM*: RBF kernel; pen. of error = 1.0; stopping criterion = 10-3
- GBDT: 100 trees; max depth = 4; learning rate = 10-2
3 Quality Measures
- Area Under the ROC Curve (AUC)
- F1
- Matthews Correlation Coefficient (MCC)
Offline Experiments: Setup(2)
Datasets (random samples)
- Training Set à 1,500 ads shown in March 2014
- Test Set à 550 ads shown in April 2014
Thresholds used for labelling
- Dwell Time (tδ) à [35,45] secs.
- Bounce Rate (τβ) à [.20,.25]
40 secs. is the median dwell
time in our training set
Offline Evaluation
P(dwell time > tδ) P(bounce rate < τβ)
All the learning methods perform
similarly
Similarity features do not improve
the performance (in fact, they can
demote it)
Content features alone achieve
good performance
Historic features boost
performance but they can be
sparse
Feature Ranking
Top-15 ranked features using GBDT classifier
No similarity feature in the top-15 when GBDT is trained on C-S-H
(first and only one is ranked 11th for C-S)
Content features are crucial
Historic features are the top-3 ranked for C-S-H
Online Bucket Testing
A/B Testing
Baseline Bucket (control) High-Quality Bucket (treatment)
Ø  LogReg implementation in
production
Ø  P(dwell time > tδ) [tδ = 40 secs.]
Ø  Only for Yahoo iOS App (mobile)
Online Experiments: Setup
2 Datasets (random samples)
- Time range: May-June 2014
- D1: contains clicks on ads served by the baseline bucket
- D2: contains clicks on ads served by the high-quality bucket
3 Analyses
- (ad-)click-level
- ad-level
- user-level
-  All à ads/users appearing in
D1 or D2
-  Shared à ads/users appearing
in D1 and D2
-  Unique à ads/users appearing
in D1 xor D2
(Ad-)Click-Level Analysis
Measure the difference between CTR
as obtained from the two buckets
All
-  14,500 ads (baseline)
-  12,500 ads (high quality)
-  +18% CTR on high quality
bucket
-  Two-proportion z-test rejects
the null hypothesis at α = .01
H0: CTRbaseline = CTRhigh-quality
Ad-Level Analysis
Measure the difference between dwell time and
bounce rate as obtained from the two buckets
from ad perspective
-  Remove
-  very long clicks (i.e. > 10 mins.)
-  ads clicked less than 10 times
-  Bounce click = dwell time < 5 secs.
-  1,000 ads (baseline)
-  700 ads (high quality)
-  600 ads (shared)
-  Two-sample K-S test significant at α = .01 except
for bounce rate on shared ads
User-Level Analysis
Measure the difference between dwell time and
bounce rate as obtained from the two buckets
from user perspective
-  Remove
-  very long clicks (i.e. > 10 mins.)
-  users who clicked on less than 2 unique ads
-  Bounce click = dwell time < 5 secs.
-  16,000 users (baseline)
-  11,000 users (high quality)
-  2,700 users (shared)
-  Two-sample K-S test significant at α = .01
Conclusions
-  Factor the post-click experience into the traditional
ad ranking scheme which optimises only for CTR
-  Define 2 proxy measures of post-click ad quality:
dwell time and bounce rate
-  Design a machine-learning-based solution to predict
high post-click quality ads from the features of the
landing page (and the creative)
-  Deploy an implementation in production on Yahoo
Gemini
-  Online A/B testing shows that dwell time
significantly increases (+30%) and bounce rate
significantly decreases (-6.7%) with only small drop
of eCPM

More Related Content

What's hot

eMarketer Webinar: Keys to Online Display Advertising
eMarketer Webinar: Keys to Online Display AdvertisingeMarketer Webinar: Keys to Online Display Advertising
eMarketer Webinar: Keys to Online Display Advertising
eMarketer
 
Behavioral targeting
Behavioral targetingBehavioral targeting
Behavioral targeting
Anil Batra
 
Digital Marketing Strategy - Canvas for Bruxelles formation training
Digital Marketing Strategy - Canvas for Bruxelles formation trainingDigital Marketing Strategy - Canvas for Bruxelles formation training
Digital Marketing Strategy - Canvas for Bruxelles formation training
David Hachez
 
Social media and reviews are changing the landscape of revenue management
Social media and reviews are changing the landscape of revenue managementSocial media and reviews are changing the landscape of revenue management
Social media and reviews are changing the landscape of revenue management
Kevin May
 
Search 101
Search 101Search 101
Search 101
The Media Kitchen
 
Social Advertising: Best Practice for Facebook Ads
Social Advertising: Best Practice for Facebook AdsSocial Advertising: Best Practice for Facebook Ads
Social Advertising: Best Practice for Facebook Ads
UMGD Direct Marketing
 
Yelp Product Challenge
Yelp Product ChallengeYelp Product Challenge
Yelp Product Challenge
Hisham Radwan
 
Digital plan moms company Delhi NCR
Digital plan moms company Delhi NCRDigital plan moms company Delhi NCR
Digital plan moms company Delhi NCR
mediaant
 
Digital display advertising through Audience Extension platform
Digital display advertising through Audience Extension platformDigital display advertising through Audience Extension platform
Digital display advertising through Audience Extension platform
Jeroen Corver
 
KREATIO-WHITE-PAPER----AD-REVENUE.PDF
KREATIO-WHITE-PAPER----AD-REVENUE.PDFKREATIO-WHITE-PAPER----AD-REVENUE.PDF
KREATIO-WHITE-PAPER----AD-REVENUE.PDF
Kadam Vivek
 
Display Advertising Technology
Display Advertising TechnologyDisplay Advertising Technology
Display Advertising Technology
Tony Passey
 
Search Engine
Search EngineSearch Engine
Search Engine
student
 
Tips & tricks how to maximise your Facebook campaign by Andrei Ursuleanu @ Al...
Tips & tricks how to maximise your Facebook campaign by Andrei Ursuleanu @ Al...Tips & tricks how to maximise your Facebook campaign by Andrei Ursuleanu @ Al...
Tips & tricks how to maximise your Facebook campaign by Andrei Ursuleanu @ Al...
ICEEFEST2013
 
Audience Extension Survey 2014
Audience Extension Survey 2014Audience Extension Survey 2014
Audience Extension Survey 2014
Jinenne Sutherland
 
Computational Advertising
Computational AdvertisingComputational Advertising
Computational Advertising
AhmadShah Sultani
 
Tutorial 10 (computational advertising)
Tutorial 10 (computational advertising)Tutorial 10 (computational advertising)
Tutorial 10 (computational advertising)
Kira
 
Tutorial 11 (computational advertising)
Tutorial 11 (computational advertising)Tutorial 11 (computational advertising)
Tutorial 11 (computational advertising)
Kira
 
How to Conduct UX Benchmarking
How to Conduct UX BenchmarkingHow to Conduct UX Benchmarking
How to Conduct UX Benchmarking
UserZoom
 
Fostering O2O Solution by Scenario and LBS Targeting - Nespresso Case Study
Fostering O2O Solution by Scenario and LBS Targeting - Nespresso Case StudyFostering O2O Solution by Scenario and LBS Targeting - Nespresso Case Study
Fostering O2O Solution by Scenario and LBS Targeting - Nespresso Case Study
vponmkt
 

What's hot (19)

eMarketer Webinar: Keys to Online Display Advertising
eMarketer Webinar: Keys to Online Display AdvertisingeMarketer Webinar: Keys to Online Display Advertising
eMarketer Webinar: Keys to Online Display Advertising
 
Behavioral targeting
Behavioral targetingBehavioral targeting
Behavioral targeting
 
Digital Marketing Strategy - Canvas for Bruxelles formation training
Digital Marketing Strategy - Canvas for Bruxelles formation trainingDigital Marketing Strategy - Canvas for Bruxelles formation training
Digital Marketing Strategy - Canvas for Bruxelles formation training
 
Social media and reviews are changing the landscape of revenue management
Social media and reviews are changing the landscape of revenue managementSocial media and reviews are changing the landscape of revenue management
Social media and reviews are changing the landscape of revenue management
 
Search 101
Search 101Search 101
Search 101
 
Social Advertising: Best Practice for Facebook Ads
Social Advertising: Best Practice for Facebook AdsSocial Advertising: Best Practice for Facebook Ads
Social Advertising: Best Practice for Facebook Ads
 
Yelp Product Challenge
Yelp Product ChallengeYelp Product Challenge
Yelp Product Challenge
 
Digital plan moms company Delhi NCR
Digital plan moms company Delhi NCRDigital plan moms company Delhi NCR
Digital plan moms company Delhi NCR
 
Digital display advertising through Audience Extension platform
Digital display advertising through Audience Extension platformDigital display advertising through Audience Extension platform
Digital display advertising through Audience Extension platform
 
KREATIO-WHITE-PAPER----AD-REVENUE.PDF
KREATIO-WHITE-PAPER----AD-REVENUE.PDFKREATIO-WHITE-PAPER----AD-REVENUE.PDF
KREATIO-WHITE-PAPER----AD-REVENUE.PDF
 
Display Advertising Technology
Display Advertising TechnologyDisplay Advertising Technology
Display Advertising Technology
 
Search Engine
Search EngineSearch Engine
Search Engine
 
Tips & tricks how to maximise your Facebook campaign by Andrei Ursuleanu @ Al...
Tips & tricks how to maximise your Facebook campaign by Andrei Ursuleanu @ Al...Tips & tricks how to maximise your Facebook campaign by Andrei Ursuleanu @ Al...
Tips & tricks how to maximise your Facebook campaign by Andrei Ursuleanu @ Al...
 
Audience Extension Survey 2014
Audience Extension Survey 2014Audience Extension Survey 2014
Audience Extension Survey 2014
 
Computational Advertising
Computational AdvertisingComputational Advertising
Computational Advertising
 
Tutorial 10 (computational advertising)
Tutorial 10 (computational advertising)Tutorial 10 (computational advertising)
Tutorial 10 (computational advertising)
 
Tutorial 11 (computational advertising)
Tutorial 11 (computational advertising)Tutorial 11 (computational advertising)
Tutorial 11 (computational advertising)
 
How to Conduct UX Benchmarking
How to Conduct UX BenchmarkingHow to Conduct UX Benchmarking
How to Conduct UX Benchmarking
 
Fostering O2O Solution by Scenario and LBS Targeting - Nespresso Case Study
Fostering O2O Solution by Scenario and LBS Targeting - Nespresso Case StudyFostering O2O Solution by Scenario and LBS Targeting - Nespresso Case Study
Fostering O2O Solution by Scenario and LBS Targeting - Nespresso Case Study
 

Similar to Promoting Positive Post-click Experience for In-Stream Yahoo Gemini Users

Alenty
AlentyAlenty
Alenty
emerceron
 
Vpon - 廣告效果導向為基礎的行動廣告系統
Vpon - 廣告效果導向為基礎的行動廣告系統Vpon - 廣告效果導向為基礎的行動廣告系統
Vpon - 廣告效果導向為基礎的行動廣告系統
Vpon
 
Alenty
AlentyAlenty
Alenty
emerceron
 
Alenty
AlentyAlenty
Alenty
emerceron
 
Data Science at Flurry
Data Science at FlurryData Science at Flurry
Data Science at Flurry
soupsranjan
 
Introduction to Alenty
Introduction to AlentyIntroduction to Alenty
Introduction to Alenty
emerceron
 
Computational Advertising in Yelp Local Ads
Computational Advertising in Yelp Local AdsComputational Advertising in Yelp Local Ads
Computational Advertising in Yelp Local Ads
soupsranjan
 
Alenty Light Master Presentation
Alenty Light   Master PresentationAlenty Light   Master Presentation
Alenty Light Master Presentation
emerceron
 
Science of retargeting
Science of retargetingScience of retargeting
Science of retargeting
Bhaskar Jayaraman
 
Presentación de Peter Minnium en IAB Conecta 2013
Presentación de Peter Minnium en IAB Conecta 2013Presentación de Peter Minnium en IAB Conecta 2013
Presentación de Peter Minnium en IAB Conecta 2013
IAB México
 
Digital analytics lecture4
Digital analytics lecture4Digital analytics lecture4
Digital analytics lecture4
Joni Salminen
 
Search Engine Advertising
Search Engine AdvertisingSearch Engine Advertising
Search Engine Advertising
Altex Marketing OÜ
 
Alenty
AlentyAlenty
Concern Digital Training 26th August 2015
Concern Digital Training 26th August 2015Concern Digital Training 26th August 2015
Concern Digital Training 26th August 2015
Vanessa Vallejo
 
What Is Google Ads?
What Is Google Ads?What Is Google Ads?
What Is Google Ads?
BIRVASONI
 
Alenty appnexus app
Alenty  appnexus appAlenty  appnexus app
Alenty appnexus app
Frank Durousset
 
Alenty appnexus app
Alenty  appnexus appAlenty  appnexus app
Alenty appnexus app
Frank Durousset
 
Alenty - leader in ad viewability measurement
Alenty - leader in ad viewability measurementAlenty - leader in ad viewability measurement
Alenty - leader in ad viewability measurement
Kévin Arnoult
 
Media planning
Media planningMedia planning
Media planning
Rbk Asr
 
1 media planning
1 media planning1 media planning
1 media planning
Rbk Asr
 

Similar to Promoting Positive Post-click Experience for In-Stream Yahoo Gemini Users (20)

Alenty
AlentyAlenty
Alenty
 
Vpon - 廣告效果導向為基礎的行動廣告系統
Vpon - 廣告效果導向為基礎的行動廣告系統Vpon - 廣告效果導向為基礎的行動廣告系統
Vpon - 廣告效果導向為基礎的行動廣告系統
 
Alenty
AlentyAlenty
Alenty
 
Alenty
AlentyAlenty
Alenty
 
Data Science at Flurry
Data Science at FlurryData Science at Flurry
Data Science at Flurry
 
Introduction to Alenty
Introduction to AlentyIntroduction to Alenty
Introduction to Alenty
 
Computational Advertising in Yelp Local Ads
Computational Advertising in Yelp Local AdsComputational Advertising in Yelp Local Ads
Computational Advertising in Yelp Local Ads
 
Alenty Light Master Presentation
Alenty Light   Master PresentationAlenty Light   Master Presentation
Alenty Light Master Presentation
 
Science of retargeting
Science of retargetingScience of retargeting
Science of retargeting
 
Presentación de Peter Minnium en IAB Conecta 2013
Presentación de Peter Minnium en IAB Conecta 2013Presentación de Peter Minnium en IAB Conecta 2013
Presentación de Peter Minnium en IAB Conecta 2013
 
Digital analytics lecture4
Digital analytics lecture4Digital analytics lecture4
Digital analytics lecture4
 
Search Engine Advertising
Search Engine AdvertisingSearch Engine Advertising
Search Engine Advertising
 
Alenty
AlentyAlenty
Alenty
 
Concern Digital Training 26th August 2015
Concern Digital Training 26th August 2015Concern Digital Training 26th August 2015
Concern Digital Training 26th August 2015
 
What Is Google Ads?
What Is Google Ads?What Is Google Ads?
What Is Google Ads?
 
Alenty appnexus app
Alenty  appnexus appAlenty  appnexus app
Alenty appnexus app
 
Alenty appnexus app
Alenty  appnexus appAlenty  appnexus app
Alenty appnexus app
 
Alenty - leader in ad viewability measurement
Alenty - leader in ad viewability measurementAlenty - leader in ad viewability measurement
Alenty - leader in ad viewability measurement
 
Media planning
Media planningMedia planning
Media planning
 
1 media planning
1 media planning1 media planning
1 media planning
 

More from Mounia Lalmas-Roelleke

Engagement, Metrics & Personalisation at Scale
Engagement, Metrics &  Personalisation at ScaleEngagement, Metrics &  Personalisation at Scale
Engagement, Metrics & Personalisation at Scale
Mounia Lalmas-Roelleke
 
Engagement, metrics and "recommenders"
Engagement, metrics and "recommenders"Engagement, metrics and "recommenders"
Engagement, metrics and "recommenders"
Mounia Lalmas-Roelleke
 
Metrics, Engagement & Personalization
Metrics, Engagement & Personalization Metrics, Engagement & Personalization
Metrics, Engagement & Personalization
Mounia Lalmas-Roelleke
 
Tutorial on Online User Engagement: Metrics and Optimization
Tutorial on Online User Engagement: Metrics and OptimizationTutorial on Online User Engagement: Metrics and Optimization
Tutorial on Online User Engagement: Metrics and Optimization
Mounia Lalmas-Roelleke
 
Recommending and searching @ Spotify
Recommending and searching @ SpotifyRecommending and searching @ Spotify
Recommending and searching @ Spotify
Mounia Lalmas-Roelleke
 
Personalizing the listening experience
Personalizing the listening experiencePersonalizing the listening experience
Personalizing the listening experience
Mounia Lalmas-Roelleke
 
Recommending and Searching (Research @ Spotify)
Recommending and Searching (Research @ Spotify)Recommending and Searching (Research @ Spotify)
Recommending and Searching (Research @ Spotify)
Mounia Lalmas-Roelleke
 
Search @ Spotify
Search @ Spotify Search @ Spotify
Search @ Spotify
Mounia Lalmas-Roelleke
 
Tutorial on metrics of user engagement -- Applications to Search & E- commerce
Tutorial on metrics of user engagement -- Applications to Search & E- commerceTutorial on metrics of user engagement -- Applications to Search & E- commerce
Tutorial on metrics of user engagement -- Applications to Search & E- commerce
Mounia Lalmas-Roelleke
 
An introduction to system-oriented evaluation in Information Retrieval
An introduction to system-oriented evaluation in Information RetrievalAn introduction to system-oriented evaluation in Information Retrieval
An introduction to system-oriented evaluation in Information Retrieval
Mounia Lalmas-Roelleke
 
Social Media and AI: Don’t forget the users
Social Media and AI: Don’t forget the usersSocial Media and AI: Don’t forget the users
Social Media and AI: Don’t forget the users
Mounia Lalmas-Roelleke
 
Describing Patterns and Disruptions in Large Scale Mobile App Usage Data
Describing Patterns and Disruptions in Large Scale Mobile App Usage DataDescribing Patterns and Disruptions in Large Scale Mobile App Usage Data
Describing Patterns and Disruptions in Large Scale Mobile App Usage Data
Mounia Lalmas-Roelleke
 
Evaluating the search experience: from Retrieval Effectiveness to User Engage...
Evaluating the search experience: from Retrieval Effectiveness to User Engage...Evaluating the search experience: from Retrieval Effectiveness to User Engage...
Evaluating the search experience: from Retrieval Effectiveness to User Engage...
Mounia Lalmas-Roelleke
 
From “Selena Gomez” to “Marlon Brando”: Understanding Explorative Entity Search
From “Selena Gomez” to “Marlon Brando”: Understanding Explorative Entity SearchFrom “Selena Gomez” to “Marlon Brando”: Understanding Explorative Entity Search
From “Selena Gomez” to “Marlon Brando”: Understanding Explorative Entity Search
Mounia Lalmas-Roelleke
 
How Big Data is Changing User Engagement
How Big Data is Changing User EngagementHow Big Data is Changing User Engagement
How Big Data is Changing User Engagement
Mounia Lalmas-Roelleke
 
Measuring user engagement: the do, the do not do, and the we do not know
Measuring user engagement: the do, the do not do, and the we do not knowMeasuring user engagement: the do, the do not do, and the we do not know
Measuring user engagement: the do, the do not do, and the we do not know
Mounia Lalmas-Roelleke
 
An Engaging Click ... or how can user engagement measurement inform web searc...
An Engaging Click ... or how can user engagement measurement inform web searc...An Engaging Click ... or how can user engagement measurement inform web searc...
An Engaging Click ... or how can user engagement measurement inform web searc...
Mounia Lalmas-Roelleke
 
Social Media News Communities: Gatekeeping, Coverage, and Statement Bias
 Social Media News Communities: Gatekeeping, Coverage, and Statement Bias Social Media News Communities: Gatekeeping, Coverage, and Statement Bias
Social Media News Communities: Gatekeeping, Coverage, and Statement Bias
Mounia Lalmas-Roelleke
 
On the Reliability and Intuitiveness of Aggregated Search Metrics
On the Reliability and Intuitiveness of Aggregated Search MetricsOn the Reliability and Intuitiveness of Aggregated Search Metrics
On the Reliability and Intuitiveness of Aggregated Search Metrics
Mounia Lalmas-Roelleke
 
Penguins in Sweaters, or Serendipitous Entity Search on User-generated Content
 Penguins in Sweaters, or Serendipitous Entity Search on User-generated Content Penguins in Sweaters, or Serendipitous Entity Search on User-generated Content
Penguins in Sweaters, or Serendipitous Entity Search on User-generated Content
Mounia Lalmas-Roelleke
 

More from Mounia Lalmas-Roelleke (20)

Engagement, Metrics & Personalisation at Scale
Engagement, Metrics &  Personalisation at ScaleEngagement, Metrics &  Personalisation at Scale
Engagement, Metrics & Personalisation at Scale
 
Engagement, metrics and "recommenders"
Engagement, metrics and "recommenders"Engagement, metrics and "recommenders"
Engagement, metrics and "recommenders"
 
Metrics, Engagement & Personalization
Metrics, Engagement & Personalization Metrics, Engagement & Personalization
Metrics, Engagement & Personalization
 
Tutorial on Online User Engagement: Metrics and Optimization
Tutorial on Online User Engagement: Metrics and OptimizationTutorial on Online User Engagement: Metrics and Optimization
Tutorial on Online User Engagement: Metrics and Optimization
 
Recommending and searching @ Spotify
Recommending and searching @ SpotifyRecommending and searching @ Spotify
Recommending and searching @ Spotify
 
Personalizing the listening experience
Personalizing the listening experiencePersonalizing the listening experience
Personalizing the listening experience
 
Recommending and Searching (Research @ Spotify)
Recommending and Searching (Research @ Spotify)Recommending and Searching (Research @ Spotify)
Recommending and Searching (Research @ Spotify)
 
Search @ Spotify
Search @ Spotify Search @ Spotify
Search @ Spotify
 
Tutorial on metrics of user engagement -- Applications to Search & E- commerce
Tutorial on metrics of user engagement -- Applications to Search & E- commerceTutorial on metrics of user engagement -- Applications to Search & E- commerce
Tutorial on metrics of user engagement -- Applications to Search & E- commerce
 
An introduction to system-oriented evaluation in Information Retrieval
An introduction to system-oriented evaluation in Information RetrievalAn introduction to system-oriented evaluation in Information Retrieval
An introduction to system-oriented evaluation in Information Retrieval
 
Social Media and AI: Don’t forget the users
Social Media and AI: Don’t forget the usersSocial Media and AI: Don’t forget the users
Social Media and AI: Don’t forget the users
 
Describing Patterns and Disruptions in Large Scale Mobile App Usage Data
Describing Patterns and Disruptions in Large Scale Mobile App Usage DataDescribing Patterns and Disruptions in Large Scale Mobile App Usage Data
Describing Patterns and Disruptions in Large Scale Mobile App Usage Data
 
Evaluating the search experience: from Retrieval Effectiveness to User Engage...
Evaluating the search experience: from Retrieval Effectiveness to User Engage...Evaluating the search experience: from Retrieval Effectiveness to User Engage...
Evaluating the search experience: from Retrieval Effectiveness to User Engage...
 
From “Selena Gomez” to “Marlon Brando”: Understanding Explorative Entity Search
From “Selena Gomez” to “Marlon Brando”: Understanding Explorative Entity SearchFrom “Selena Gomez” to “Marlon Brando”: Understanding Explorative Entity Search
From “Selena Gomez” to “Marlon Brando”: Understanding Explorative Entity Search
 
How Big Data is Changing User Engagement
How Big Data is Changing User EngagementHow Big Data is Changing User Engagement
How Big Data is Changing User Engagement
 
Measuring user engagement: the do, the do not do, and the we do not know
Measuring user engagement: the do, the do not do, and the we do not knowMeasuring user engagement: the do, the do not do, and the we do not know
Measuring user engagement: the do, the do not do, and the we do not know
 
An Engaging Click ... or how can user engagement measurement inform web searc...
An Engaging Click ... or how can user engagement measurement inform web searc...An Engaging Click ... or how can user engagement measurement inform web searc...
An Engaging Click ... or how can user engagement measurement inform web searc...
 
Social Media News Communities: Gatekeeping, Coverage, and Statement Bias
 Social Media News Communities: Gatekeeping, Coverage, and Statement Bias Social Media News Communities: Gatekeeping, Coverage, and Statement Bias
Social Media News Communities: Gatekeeping, Coverage, and Statement Bias
 
On the Reliability and Intuitiveness of Aggregated Search Metrics
On the Reliability and Intuitiveness of Aggregated Search MetricsOn the Reliability and Intuitiveness of Aggregated Search Metrics
On the Reliability and Intuitiveness of Aggregated Search Metrics
 
Penguins in Sweaters, or Serendipitous Entity Search on User-generated Content
 Penguins in Sweaters, or Serendipitous Entity Search on User-generated Content Penguins in Sweaters, or Serendipitous Entity Search on User-generated Content
Penguins in Sweaters, or Serendipitous Entity Search on User-generated Content
 

Recently uploaded

[HUN][hackersuli] Red Teaming alapok 2024
[HUN][hackersuli] Red Teaming alapok 2024[HUN][hackersuli] Red Teaming alapok 2024
[HUN][hackersuli] Red Teaming alapok 2024
hackersuli
 
不能毕业如何获得(USYD毕业证)悉尼大学毕业证成绩单一比一原版制作
不能毕业如何获得(USYD毕业证)悉尼大学毕业证成绩单一比一原版制作不能毕业如何获得(USYD毕业证)悉尼大学毕业证成绩单一比一原版制作
不能毕业如何获得(USYD毕业证)悉尼大学毕业证成绩单一比一原版制作
bseovas
 
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaalmanuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
wolfsoftcompanyco
 
Should Repositories Participate in the Fediverse?
Should Repositories Participate in the Fediverse?Should Repositories Participate in the Fediverse?
Should Repositories Participate in the Fediverse?
Paul Walk
 
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
ysasp1
 
办理新西兰奥克兰大学毕业证学位证书范本原版一模一样
办理新西兰奥克兰大学毕业证学位证书范本原版一模一样办理新西兰奥克兰大学毕业证学位证书范本原版一模一样
办理新西兰奥克兰大学毕业证学位证书范本原版一模一样
xjq03c34
 
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
uehowe
 
Gen Z and the marketplaces - let's translate their needs
Gen Z and the marketplaces - let's translate their needsGen Z and the marketplaces - let's translate their needs
Gen Z and the marketplaces - let's translate their needs
Laura Szabó
 
制作毕业证书(ANU毕业证)莫纳什大学毕业证成绩单官方原版办理
制作毕业证书(ANU毕业证)莫纳什大学毕业证成绩单官方原版办理制作毕业证书(ANU毕业证)莫纳什大学毕业证成绩单官方原版办理
制作毕业证书(ANU毕业证)莫纳什大学毕业证成绩单官方原版办理
cuobya
 
留学学历(UoA毕业证)奥克兰大学毕业证成绩单官方原版办理
留学学历(UoA毕业证)奥克兰大学毕业证成绩单官方原版办理留学学历(UoA毕业证)奥克兰大学毕业证成绩单官方原版办理
留学学历(UoA毕业证)奥克兰大学毕业证成绩单官方原版办理
bseovas
 
HijackLoader Evolution: Interactive Process Hollowing
HijackLoader Evolution: Interactive Process HollowingHijackLoader Evolution: Interactive Process Hollowing
HijackLoader Evolution: Interactive Process Hollowing
Donato Onofri
 
Discover the benefits of outsourcing SEO to India
Discover the benefits of outsourcing SEO to IndiaDiscover the benefits of outsourcing SEO to India
Discover the benefits of outsourcing SEO to India
davidjhones387
 
快速办理(新加坡SMU毕业证书)新加坡管理大学毕业证文凭证书一模一样
快速办理(新加坡SMU毕业证书)新加坡管理大学毕业证文凭证书一模一样快速办理(新加坡SMU毕业证书)新加坡管理大学毕业证文凭证书一模一样
快速办理(新加坡SMU毕业证书)新加坡管理大学毕业证文凭证书一模一样
3a0sd7z3
 
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
zoowe
 
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
uehowe
 
Search Result Showing My Post is Now Buried
Search Result Showing My Post is Now BuriedSearch Result Showing My Post is Now Buried
Search Result Showing My Post is Now Buried
Trish Parr
 
一比一原版(USYD毕业证)悉尼大学毕业证如何办理
一比一原版(USYD毕业证)悉尼大学毕业证如何办理一比一原版(USYD毕业证)悉尼大学毕业证如何办理
一比一原版(USYD毕业证)悉尼大学毕业证如何办理
k4ncd0z
 
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
vmemo1
 
Azure EA Sponsorship - Customer Guide.pdf
Azure EA Sponsorship - Customer Guide.pdfAzure EA Sponsorship - Customer Guide.pdf
Azure EA Sponsorship - Customer Guide.pdf
AanSulistiyo
 
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
fovkoyb
 

Recently uploaded (20)

[HUN][hackersuli] Red Teaming alapok 2024
[HUN][hackersuli] Red Teaming alapok 2024[HUN][hackersuli] Red Teaming alapok 2024
[HUN][hackersuli] Red Teaming alapok 2024
 
不能毕业如何获得(USYD毕业证)悉尼大学毕业证成绩单一比一原版制作
不能毕业如何获得(USYD毕业证)悉尼大学毕业证成绩单一比一原版制作不能毕业如何获得(USYD毕业证)悉尼大学毕业证成绩单一比一原版制作
不能毕业如何获得(USYD毕业证)悉尼大学毕业证成绩单一比一原版制作
 
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaalmanuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
manuaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaal
 
Should Repositories Participate in the Fediverse?
Should Repositories Participate in the Fediverse?Should Repositories Participate in the Fediverse?
Should Repositories Participate in the Fediverse?
 
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
成绩单ps(UST毕业证)圣托马斯大学毕业证成绩单快速办理
 
办理新西兰奥克兰大学毕业证学位证书范本原版一模一样
办理新西兰奥克兰大学毕业证学位证书范本原版一模一样办理新西兰奥克兰大学毕业证学位证书范本原版一模一样
办理新西兰奥克兰大学毕业证学位证书范本原版一模一样
 
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
办理毕业证(NYU毕业证)纽约大学毕业证成绩单官方原版办理
 
Gen Z and the marketplaces - let's translate their needs
Gen Z and the marketplaces - let's translate their needsGen Z and the marketplaces - let's translate their needs
Gen Z and the marketplaces - let's translate their needs
 
制作毕业证书(ANU毕业证)莫纳什大学毕业证成绩单官方原版办理
制作毕业证书(ANU毕业证)莫纳什大学毕业证成绩单官方原版办理制作毕业证书(ANU毕业证)莫纳什大学毕业证成绩单官方原版办理
制作毕业证书(ANU毕业证)莫纳什大学毕业证成绩单官方原版办理
 
留学学历(UoA毕业证)奥克兰大学毕业证成绩单官方原版办理
留学学历(UoA毕业证)奥克兰大学毕业证成绩单官方原版办理留学学历(UoA毕业证)奥克兰大学毕业证成绩单官方原版办理
留学学历(UoA毕业证)奥克兰大学毕业证成绩单官方原版办理
 
HijackLoader Evolution: Interactive Process Hollowing
HijackLoader Evolution: Interactive Process HollowingHijackLoader Evolution: Interactive Process Hollowing
HijackLoader Evolution: Interactive Process Hollowing
 
Discover the benefits of outsourcing SEO to India
Discover the benefits of outsourcing SEO to IndiaDiscover the benefits of outsourcing SEO to India
Discover the benefits of outsourcing SEO to India
 
快速办理(新加坡SMU毕业证书)新加坡管理大学毕业证文凭证书一模一样
快速办理(新加坡SMU毕业证书)新加坡管理大学毕业证文凭证书一模一样快速办理(新加坡SMU毕业证书)新加坡管理大学毕业证文凭证书一模一样
快速办理(新加坡SMU毕业证书)新加坡管理大学毕业证文凭证书一模一样
 
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
国外证书(Lincoln毕业证)新西兰林肯大学毕业证成绩单不能毕业办理
 
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
留学挂科(UofM毕业证)明尼苏达大学毕业证成绩单复刻办理
 
Search Result Showing My Post is Now Buried
Search Result Showing My Post is Now BuriedSearch Result Showing My Post is Now Buried
Search Result Showing My Post is Now Buried
 
一比一原版(USYD毕业证)悉尼大学毕业证如何办理
一比一原版(USYD毕业证)悉尼大学毕业证如何办理一比一原版(USYD毕业证)悉尼大学毕业证如何办理
一比一原版(USYD毕业证)悉尼大学毕业证如何办理
 
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
重新申请毕业证书(RMIT毕业证)皇家墨尔本理工大学毕业证成绩单精仿办理
 
Azure EA Sponsorship - Customer Guide.pdf
Azure EA Sponsorship - Customer Guide.pdfAzure EA Sponsorship - Customer Guide.pdf
Azure EA Sponsorship - Customer Guide.pdf
 
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
存档可查的(USC毕业证)南加利福尼亚大学毕业证成绩单制做办理
 

Promoting Positive Post-click Experience for In-Stream Yahoo Gemini Users

  • 1. Promoting Positive Post-Click Experience for In-Stream Yahoo Gemini Users Janette Lehmann Freie Universität Berlin Berlin, Germany Mounia Lalmas Yahoo Labs London, UK Guy Shaked Yahoo Tel Aviv, Israel Gabriele Tolomei Yahoo Labs London, UK Fabrizio Silvestri Yahoo Labs London, UK
  • 2. The Online Advertising Arena PUBLISHER Maximise revenue from online advertising
  • 3. The Online Advertising Arena ADVERTISER Maximise ROI from his ad campaigns PUBLISHER Maximise revenue from online advertising
  • 4. The Online Advertising Arena ADVERTISER Maximise ROI from his ad campaigns PUBLISHER Maximise revenue from online advertising USER Show me some content, not these #*!%ing ads (… unless they are “interesting”)
  • 5. The Online Advertising Arena We want of course to maximise our revenue and guaranteeing ROI to our advertisers without negatively impacting on user experience
  • 6. Native Ads on Yahoo Stream Gemini is the Yahoo’s marketplace for native ads (mobile + desktop) and sponsored ads (mobile) …
  • 7. Native Ads on Yahoo Stream Gemini is the Yahoo’s marketplace for native ads (mobile + desktop) and sponsored ads (mobile)
  • 8. Based on eCPM (expected Cost Per Mille Impressions) CPC (Cost Per Click) CPC is the amount of money (bid) the advertiser is willing to pay for 1 click on its ad eCPM is computed from the underlying pricing model rank(ad) = eCPM(ad) Ad Ranking rank(ad) = CPC(ad) * eCTR(ad) [* 1000] P(click|ad)
  • 9. Beyond CTR Optimisation Ranking via CTR optimisation is an attempt to maximise revenue on the short-term Trade-off between provider’s and advertiser’s need What about the user? Better to include a “quality” factor so to optimise for ads having “high quality” rank(ad) = CPC(ad) * eCTR(ad) P(click|ad) P(hq|click) * eHQ(ad)
  • 10. How Do We Estimate Quality? “High Quality” is a shorthand for “High Quality Post- Click User Experience” Intuitively a high quality ad is one whose conversion rate is high Estimating the probability of conversion is hard! - Conversion data is unavailable for a large set of advertisers - Conversion data is not missing at random Use “proxy” metrics of post-click quality like dwell time and bounce rate on the ad landing page
  • 11. Post-Click Quality Metrics tad-click tback-to-stream dwell time = tback-to-stream – tad-click bounce rate = P(dwell time < th) th = 5 secs. (mobile) 12 secs. (desktop)
  • 12. Is Dwell Time A Good Proxy? mobile 200K ad clicks Ø  It needs less time to get the same probability of a second click desktop (Yahoo toolbar) 30K ad clicks Ø  23.3% of users visit other websites than the ad landing page before returning to Yahoo Ø  this goes down to 7.4% for dwell time up to 3 mins. Probability of a second click increases with dwell time
  • 13. Optimise for High Quality Ads Estimating P(hq|click) turns into estimating: P(dwell time > tδ) P(bounce rate < τβ) Can we build a model which is able to predict if an ad is high quality according to the definitions above? è  high quality = high dwell time è high quality = low bounce rate or rank(ad) = CPC(ad) * eCTR(ad) P(click|ad) P(hq|click) * eHQ(ad)
  • 14. Binary Classification Problem AD 1 2 … i … m ad landing page Xi class label Yi +1 if hq(i) -1 o.w. dwell time > tδ or bounce rate < τβ f(Xi) Feature-based representation
  • 15. Binary Classification Problem find the estimate which best approximates P(Yi | f(Xi))
  • 16. Features How do we represent an ad landing page Xi? CONTENT clickToCall numClickable tokenCount numImages … isMobileOptimised SIMILARITY cosineCreativeLandingPageNouns cosineCreativeLandingPageWiki HISTORICAL numImpressions numClicks bouncerate avgDwellTime avgDwellTimeNonShort ctr cpc
  • 17. Offline Experiments: Setup(1) *Class probability obtained using Platt scaling 3 Learning Methods - LogReg: L1-norm; C=100; stopping criterion = 10-2 - SVM*: RBF kernel; pen. of error = 1.0; stopping criterion = 10-3 - GBDT: 100 trees; max depth = 4; learning rate = 10-2 3 Quality Measures - Area Under the ROC Curve (AUC) - F1 - Matthews Correlation Coefficient (MCC)
  • 18. Offline Experiments: Setup(2) Datasets (random samples) - Training Set à 1,500 ads shown in March 2014 - Test Set à 550 ads shown in April 2014 Thresholds used for labelling - Dwell Time (tδ) à [35,45] secs. - Bounce Rate (τβ) à [.20,.25] 40 secs. is the median dwell time in our training set
  • 19. Offline Evaluation P(dwell time > tδ) P(bounce rate < τβ) All the learning methods perform similarly Similarity features do not improve the performance (in fact, they can demote it) Content features alone achieve good performance Historic features boost performance but they can be sparse
  • 20. Feature Ranking Top-15 ranked features using GBDT classifier No similarity feature in the top-15 when GBDT is trained on C-S-H (first and only one is ranked 11th for C-S) Content features are crucial Historic features are the top-3 ranked for C-S-H
  • 21. Online Bucket Testing A/B Testing Baseline Bucket (control) High-Quality Bucket (treatment) Ø  LogReg implementation in production Ø  P(dwell time > tδ) [tδ = 40 secs.] Ø  Only for Yahoo iOS App (mobile)
  • 22. Online Experiments: Setup 2 Datasets (random samples) - Time range: May-June 2014 - D1: contains clicks on ads served by the baseline bucket - D2: contains clicks on ads served by the high-quality bucket 3 Analyses - (ad-)click-level - ad-level - user-level -  All à ads/users appearing in D1 or D2 -  Shared à ads/users appearing in D1 and D2 -  Unique à ads/users appearing in D1 xor D2
  • 23. (Ad-)Click-Level Analysis Measure the difference between CTR as obtained from the two buckets All -  14,500 ads (baseline) -  12,500 ads (high quality) -  +18% CTR on high quality bucket -  Two-proportion z-test rejects the null hypothesis at α = .01 H0: CTRbaseline = CTRhigh-quality
  • 24. Ad-Level Analysis Measure the difference between dwell time and bounce rate as obtained from the two buckets from ad perspective -  Remove -  very long clicks (i.e. > 10 mins.) -  ads clicked less than 10 times -  Bounce click = dwell time < 5 secs. -  1,000 ads (baseline) -  700 ads (high quality) -  600 ads (shared) -  Two-sample K-S test significant at α = .01 except for bounce rate on shared ads
  • 25. User-Level Analysis Measure the difference between dwell time and bounce rate as obtained from the two buckets from user perspective -  Remove -  very long clicks (i.e. > 10 mins.) -  users who clicked on less than 2 unique ads -  Bounce click = dwell time < 5 secs. -  16,000 users (baseline) -  11,000 users (high quality) -  2,700 users (shared) -  Two-sample K-S test significant at α = .01
  • 26. Conclusions -  Factor the post-click experience into the traditional ad ranking scheme which optimises only for CTR -  Define 2 proxy measures of post-click ad quality: dwell time and bounce rate -  Design a machine-learning-based solution to predict high post-click quality ads from the features of the landing page (and the creative) -  Deploy an implementation in production on Yahoo Gemini -  Online A/B testing shows that dwell time significantly increases (+30%) and bounce rate significantly decreases (-6.7%) with only small drop of eCPM