SlideShare a Scribd company logo
1 of 27
Download to read offline
LA Machine Learning Group
September 15, 2014
About IRIS.TV
Leading in-player video recommendation engine designed to create continuous streams
of personalized video
SaaS based solution for publishers and video technology companies looking to increase
views per viewing session.
IRIS.TV enables publishers to:
–  connect to their audience
–  increase video views
–  control their programming to maximize reach and revenue potential of every
viewer
How it Works
BEFORE IRIS.TV
User has to navigate back to menu or app to search for more content
How it Works
Powered by Adaptive StreamTM, IRIS.TV
•  Adaptive machine learning
•  Programmatic video delivery
•  Personalizing experiences for audiences while enforcing editorial and business rules for
publishers
•  Improving content discovery in a manner that facilitates audience growth and interaction
•  Re-inventing the way users consume video
How it Works
Viewers of IRIS.TV enabled video players
WATCH MORE, WATCH LONGER, WATCH OFTEN
AFTER IRIS.TV INTEGRATION
Video publishers are able to engage users for longer and serve more ads
Speakers
Tom Sullivan
Chief Data Scientist
Joel Spitalnik
VP Engineering
Agenda
Personalization Response Curve
Definition of Terms
Data Science
Engineering
Summary
Personalization-Value Relationship
Personalization
Value
No Response = Unemployment
Positive Correlation
Something else
Relationship Between Personalization and Engagement
Stable
Loss of value due to
“creepiness” and loss
of serendipity
Strict increase in value
(not likely given constraints)
Value
Personalization
Definitions
Asset: A good that may be represented in an online environment and consumed in
computerized or physical form
Examples: Video, News article, Image, Sound file, Coupon, Book
Track: An ordered collection of co-presented assets
Anchor Asset: The first asset presented in a track (via landing or organic search)
Present Deliver Consume
(in part or whole)Select
Definitions
Experience All tracks that result from interactions with a distinct anchor asset
Engagement A user session that begins with the first track and ends after the final asset is
presented and/or consumed
Learning is based on track/experience/engagement history
Asset	
  A	
  
(Anchor)	
  
Asset	
  D	
   Asset	
  F	
  
Asset	
  A	
  
(Anchor)	
  
Asset	
  G	
   Asset	
  I	
  
Asset	
  B	
  
(Anchor)	
  
Asset	
  Q	
   Asset	
  T	
  
Dynamic
Update
TRACK
A1
TRACK
A2
TRACK
B1
EXPERIENCEEXPERIENCE
ENGAGEMENT
Hybrid (front/back end) Personalization
IRIS.TV ApproachValue
Personalization	
  P0	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  P1	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  P2	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  P3	
  	
  	
  	
  	
  	
  	
  P4	
  	
  	
  	
  	
  	
  	
  	
  	
  P5	
  
The Constrained Optimization Problem
Provide personalized recommendations, P, that maximize the value of variable on
the vertical axis of the response curve, V (value)
Subject to constraints (including estimation risk), : Λ
While remaining aware of residual benefits, R
●  The data scientist's desire execute machine learning to improve accuracy of predictions (recommended
assets)
●  The opportunity to generate residual value based business insights
That can be computed given the available resource environment and acceptable
latency in updates to tracks, E
max (V + R) = f(P) s.t. Λ,Ε
“A sensible estimate is an interval estimate”
IRIS.TV uses the “Strategy first” approach
rather than a “data first” approach
Building Personalized Tracks
For any given anchor asset, a pre-calculated track is based on an evaluation of
pairwise asset similarities reflecting a linear combination of
An asset-to-asset structural
component: a weighted
composite similarities among
the k asset features
s(1)
A user behavior component
based on historical interactions
with presented tracks
s(2)
Dynamically
adjusted in
based on
user
interactions
with current
track
Track Construction is a Clustering Problem
Note: similarities may not necessarily be symmetric when using user behavior data
Asset j
(in track)
Present Consume Like:
Dislike
Skips Fully
Consume
Similarity
S(2)
2 10,345 95.1% 7.2 : 1 11.6% 77.3% S(2)
1,2
7 4,235 9.1% 1 : 27.9 54.3% 15.9% S(2)
1,7
Sample of Historical Data Related to Paired Asset
Engagements for Anchor Asset 1
Temporal Observation Windows
Temporal user behavior data may perish over time
In our machine learning “sandbox”, there is little need to store and process data that has “aged” beyond
some threshold
The “optimal” temporal window may be inferred by using subject matter expertise and/or by ML
We only consider the sufficient statistics, aggregated from the full log data and updated. This reduces the
amount of data in the current environment and may improve computational efficiency, E
max (V + R) = f(P) s.t. Λ, Ε
Age	
  
Obs	
  weight	
  
t*	
  
1	
  
0	
  
Preparing the data for similarity calculations
API
Aggregation Queries
Summary
observation data
in temporal
window,
constraints, asset
metadata
Raw Log
Data
The Data Scientist at IRIS.TV
Skills and Experience
●  Ability to program in R
●  Understanding of MySQL
●  Some familiarity with working with large, diverse data sets
●  Can move data between Hive/MySQL/R
Differentiator
●  Primary Strength is Statistics v. Data Management
●  Understands dimension reduction and the pros/cons of using various
methods
●  Choices are driven by objective function
max (V + R) = f(P) s.t. Λ,Ε
Join us?: www.iris.tv
Email: tom@iris.tv joel@iris.tv
Feature 1:
Keyword
Overlap
Feature 2: Asset
Length
Tracks are Sensitive to Choice of Methods
Asset A
(anchor)
Asset B Asset C
A B C
D
E
Asset A
(anchor)
Asset B Asset E
Asset A
(anchor)
Asset B Asset D
Single-link
2NN
Complete
Complete method:
•  leads to possibly fewer computations
(when n is sufficiently large)
•  Can result more spherical clusters with
respect to anchor asset
•  Increases “track strength”
Increasing Personalization: Asset-to-Asset Feature Weighting
•  Which features are more relevant in the user’s decision to consume assets?
•  Pairwise similarity, between two assets s(1)
i,j are a weighted sum of k separate feature-
level similarities from asset metadata (e.g. length, keyword overlap, genre,
publication date, mood, etc.)
•  We use simulated annealing, simplex-marching, tree-based starting points, parsimony penalty, and
loss values driven by (E,R)
Increasing Personalization with Groups
The user population may be partitioned into groups
based on observable characteristics and revealed
feature preferences, each having a distinct set of
composite similarities – leading to possibility of
different tracks for each group
Personalization-driven user partitioning based on location
(and possibly device type) to improve “cold start”
Other	
  Fans	
  
World
Asia	
  Europe	
  
S.	
  America	
  
Not	
  Brazil	
  
N.	
  America	
  
USA	
  
Not	
  USA	
  
Red	
  Sox	
  Fans	
   Yankee	
  Fans	
  Bots	
  
Brazil	
  
Bots and Bias
•  Not all user engagement data is useful – there are non-human
users whose influence we try to remove / downweight the user
behavior component of similarity computation
•  e.g. One user selected 1600 videos in 10 minutes
•  In polarizing topics such as sports, politics, etc., and quality
assessments there may be bias in the like /dislike actions
Braz	
  
USA	
  
Red	
  Sox	
  Fans	
   Yankee	
  Fans	
   Other	
  Fans	
  Bots	
  
X	
  
Bucket Re-Evaluation
Bucket assignment is a fuzzy classification problem
and we assign a user to the bucket with the highest
membership probability
Periodically, we revisit the buckets and memberships
•  If every bucket has only one user, the backend has generated
personalization in its highest form (with respect to the IRIS.TV
definition)
•  As more data becomes available on a user based on their
historical engagements, their bucket assignment may change
Hybrid (front/back-end) Personalization, so far
Value
Personalization
	
  P0	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  P1	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  P2	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  P3	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  P4	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  P5	
  
Anchoring
+similarity	
  
+	
  Historical	
  user	
  
data	
  
+	
  buckets	
  
+beVer	
  buckets	
  
+real-­‐Xme	
  
feedback	
  
Next	
  up:	
  Joel	
  Spitalnik,	
  to	
  talk	
  
about	
  the	
  technology	
  stack,	
  NLP,	
  
and	
  dynamic	
  track	
  adjustment	
  	
  
Front End
Origins of Recommendation System
JukeboxTV
Manual Curation
Building Data Structures
Data Munging
Natural Language Processing
Mapping Keywords
Service Oriented Architecture
Agile Development
Minimizing Latency
Preparing the data for similarity calculations
Online Rec
System
Aggregation Queries
Summary
observation data
in temporal
window,
constraints, asset
metadata
Raw Log
Data
API
Thank You
Questions?
www.iris.tv tom@iris.tv joel@iris.tv rohan@iris.tv

More Related Content

Viewers also liked

Proyecto estratégico. powerPOINT
Proyecto estratégico. powerPOINTProyecto estratégico. powerPOINT
Proyecto estratégico. powerPOINTFlor Florencia
 
Взаимодействие высшего образования и рынка труда: проблемы и направления разв...
Взаимодействие высшего образования и рынка труда: проблемы и направления разв...Взаимодействие высшего образования и рынка труда: проблемы и направления разв...
Взаимодействие высшего образования и рынка труда: проблемы и направления разв...FutureToday
 
здоровье 2014-2015-сайт
здоровье 2014-2015-сайтздоровье 2014-2015-сайт
здоровье 2014-2015-сайтbakhova
 
Applying UX strategy to optimising the support experience
Applying UX strategy to optimising the support experienceApplying UX strategy to optimising the support experience
Applying UX strategy to optimising the support experienceProduct Anonymous
 
Virtual Reality Product Management - November 2016- Product Anonymous
Virtual Reality Product Management - November 2016- Product AnonymousVirtual Reality Product Management - November 2016- Product Anonymous
Virtual Reality Product Management - November 2016- Product AnonymousProduct Anonymous
 
User Experience Design: The Past, The Present, The Future
User Experience Design: The Past, The Present, The FutureUser Experience Design: The Past, The Present, The Future
User Experience Design: The Past, The Present, The FutureCharbel Zeaiter
 
La pintura en la 2ª mitad del siglo xx
La pintura en la  2ª mitad del siglo xxLa pintura en la  2ª mitad del siglo xx
La pintura en la 2ª mitad del siglo xxfernando rodriguez
 

Viewers also liked (10)

Proyecto estratégico. powerPOINT
Proyecto estratégico. powerPOINTProyecto estratégico. powerPOINT
Proyecto estratégico. powerPOINT
 
Presentación
PresentaciónPresentación
Presentación
 
Взаимодействие высшего образования и рынка труда: проблемы и направления разв...
Взаимодействие высшего образования и рынка труда: проблемы и направления разв...Взаимодействие высшего образования и рынка труда: проблемы и направления разв...
Взаимодействие высшего образования и рынка труда: проблемы и направления разв...
 
Siri lappteknik kudde
Siri lappteknik kudde Siri lappteknik kudde
Siri lappteknik kudde
 
здоровье 2014-2015-сайт
здоровье 2014-2015-сайтздоровье 2014-2015-сайт
здоровье 2014-2015-сайт
 
Applying UX strategy to optimising the support experience
Applying UX strategy to optimising the support experienceApplying UX strategy to optimising the support experience
Applying UX strategy to optimising the support experience
 
Los limites del arte 2016
Los limites del arte 2016Los limites del arte 2016
Los limites del arte 2016
 
Virtual Reality Product Management - November 2016- Product Anonymous
Virtual Reality Product Management - November 2016- Product AnonymousVirtual Reality Product Management - November 2016- Product Anonymous
Virtual Reality Product Management - November 2016- Product Anonymous
 
User Experience Design: The Past, The Present, The Future
User Experience Design: The Past, The Present, The FutureUser Experience Design: The Past, The Present, The Future
User Experience Design: The Past, The Present, The Future
 
La pintura en la 2ª mitad del siglo xx
La pintura en la  2ª mitad del siglo xxLa pintura en la  2ª mitad del siglo xx
La pintura en la 2ª mitad del siglo xx
 

Similar to IRIS.TV Talks Future of Video Personalization at Cross Campus LA

Cloudera Movies Data Science Project On Big Data
Cloudera Movies Data Science Project On Big DataCloudera Movies Data Science Project On Big Data
Cloudera Movies Data Science Project On Big DataAbhishek M Shivalingaiah
 
Recommender Systems from A to Z – The Right Dataset
Recommender Systems from A to Z – The Right DatasetRecommender Systems from A to Z – The Right Dataset
Recommender Systems from A to Z – The Right DatasetCrossing Minds
 
Ilian Uzunov (Georgi Georgiev): Ilian Uzunov (Georgi Georgiev)
Ilian Uzunov (Georgi Georgiev): Ilian Uzunov (Georgi Georgiev)Ilian Uzunov (Georgi Georgiev): Ilian Uzunov (Georgi Georgiev)
Ilian Uzunov (Georgi Georgiev): Ilian Uzunov (Georgi Georgiev)Semantic Web Company
 
Recommendation systems
Recommendation systemsRecommendation systems
Recommendation systemsAnton Ermak
 
Explain Yourself: Why You Get the Recommendations You Do
Explain Yourself: Why You Get the Recommendations You DoExplain Yourself: Why You Get the Recommendations You Do
Explain Yourself: Why You Get the Recommendations You DoDatabricks
 
Monitoring modern applications using Elastic
Monitoring modern applications using ElasticMonitoring modern applications using Elastic
Monitoring modern applications using ElasticElasticsearch
 
Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Scienc...
Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Scienc...Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Scienc...
Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Scienc...Data Con LA
 
Content - Based Recommendations Enhanced with Collaborative Information
Content - Based Recommendations Enhanced with Collaborative InformationContent - Based Recommendations Enhanced with Collaborative Information
Content - Based Recommendations Enhanced with Collaborative InformationAlessandro Liparoti
 
Scaling DDS to Millions of Computers and Devices
Scaling DDS to Millions of Computers and DevicesScaling DDS to Millions of Computers and Devices
Scaling DDS to Millions of Computers and DevicesRick Warren
 
Projection Multi Scale Hashing Keyword Search in Multidimensional Datasets
Projection Multi Scale Hashing Keyword Search in Multidimensional DatasetsProjection Multi Scale Hashing Keyword Search in Multidimensional Datasets
Projection Multi Scale Hashing Keyword Search in Multidimensional DatasetsIRJET Journal
 
Your learning ecosystem
Your learning ecosystemYour learning ecosystem
Your learning ecosystemNetDimensions
 
Combining Logs, Metrics, and Traces for Unified Observability
Combining Logs, Metrics, and Traces for Unified ObservabilityCombining Logs, Metrics, and Traces for Unified Observability
Combining Logs, Metrics, and Traces for Unified ObservabilityElasticsearch
 
C19013010 the tutorial to build shared ai services session 1
C19013010  the tutorial to build shared ai services session 1C19013010  the tutorial to build shared ai services session 1
C19013010 the tutorial to build shared ai services session 1Bill Liu
 
Data-Driven Transformation: Leveraging Big Data at Showtime with Apache Spark
Data-Driven Transformation: Leveraging Big Data at Showtime with Apache SparkData-Driven Transformation: Leveraging Big Data at Showtime with Apache Spark
Data-Driven Transformation: Leveraging Big Data at Showtime with Apache SparkDatabricks
 
Combining Logs, Metrics, and Traces for Unified Observability
Combining Logs, Metrics, and Traces for Unified ObservabilityCombining Logs, Metrics, and Traces for Unified Observability
Combining Logs, Metrics, and Traces for Unified ObservabilityElasticsearch
 
Multi-modal sources for predictive modeling using deep learning
Multi-modal sources for predictive modeling using deep learningMulti-modal sources for predictive modeling using deep learning
Multi-modal sources for predictive modeling using deep learningSanghamitra Deb
 
Using Implicit Preference Relations to Improve Content-based Recommendations,...
Using Implicit Preference Relations to Improve Content-based Recommendations,...Using Implicit Preference Relations to Improve Content-based Recommendations,...
Using Implicit Preference Relations to Improve Content-based Recommendations,...Ladislav Peska
 
Feature drift monitoring as a service for machine learning models at scale
Feature drift monitoring as a service for machine learning models at scaleFeature drift monitoring as a service for machine learning models at scale
Feature drift monitoring as a service for machine learning models at scaleNoriaki Tatsumi
 

Similar to IRIS.TV Talks Future of Video Personalization at Cross Campus LA (20)

Cloudera Movies Data Science Project On Big Data
Cloudera Movies Data Science Project On Big DataCloudera Movies Data Science Project On Big Data
Cloudera Movies Data Science Project On Big Data
 
Recommender Systems from A to Z – The Right Dataset
Recommender Systems from A to Z – The Right DatasetRecommender Systems from A to Z – The Right Dataset
Recommender Systems from A to Z – The Right Dataset
 
Ilian Uzunov (Georgi Georgiev): Ilian Uzunov (Georgi Georgiev)
Ilian Uzunov (Georgi Georgiev): Ilian Uzunov (Georgi Georgiev)Ilian Uzunov (Georgi Georgiev): Ilian Uzunov (Georgi Georgiev)
Ilian Uzunov (Georgi Georgiev): Ilian Uzunov (Georgi Georgiev)
 
Recommendation systems
Recommendation systemsRecommendation systems
Recommendation systems
 
Explain Yourself: Why You Get the Recommendations You Do
Explain Yourself: Why You Get the Recommendations You DoExplain Yourself: Why You Get the Recommendations You Do
Explain Yourself: Why You Get the Recommendations You Do
 
Monitoring modern applications using Elastic
Monitoring modern applications using ElasticMonitoring modern applications using Elastic
Monitoring modern applications using Elastic
 
Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Scienc...
Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Scienc...Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Scienc...
Big Data Day LA 2016/ Use Case Driven track - Shaping the Role of Data Scienc...
 
Content - Based Recommendations Enhanced with Collaborative Information
Content - Based Recommendations Enhanced with Collaborative InformationContent - Based Recommendations Enhanced with Collaborative Information
Content - Based Recommendations Enhanced with Collaborative Information
 
Scaling DDS to Millions of Computers and Devices
Scaling DDS to Millions of Computers and DevicesScaling DDS to Millions of Computers and Devices
Scaling DDS to Millions of Computers and Devices
 
Projection Multi Scale Hashing Keyword Search in Multidimensional Datasets
Projection Multi Scale Hashing Keyword Search in Multidimensional DatasetsProjection Multi Scale Hashing Keyword Search in Multidimensional Datasets
Projection Multi Scale Hashing Keyword Search in Multidimensional Datasets
 
kdd2015
kdd2015kdd2015
kdd2015
 
Your learning ecosystem
Your learning ecosystemYour learning ecosystem
Your learning ecosystem
 
Combining Logs, Metrics, and Traces for Unified Observability
Combining Logs, Metrics, and Traces for Unified ObservabilityCombining Logs, Metrics, and Traces for Unified Observability
Combining Logs, Metrics, and Traces for Unified Observability
 
C19013010 the tutorial to build shared ai services session 1
C19013010  the tutorial to build shared ai services session 1C19013010  the tutorial to build shared ai services session 1
C19013010 the tutorial to build shared ai services session 1
 
Data-Driven Transformation: Leveraging Big Data at Showtime with Apache Spark
Data-Driven Transformation: Leveraging Big Data at Showtime with Apache SparkData-Driven Transformation: Leveraging Big Data at Showtime with Apache Spark
Data-Driven Transformation: Leveraging Big Data at Showtime with Apache Spark
 
Combining Logs, Metrics, and Traces for Unified Observability
Combining Logs, Metrics, and Traces for Unified ObservabilityCombining Logs, Metrics, and Traces for Unified Observability
Combining Logs, Metrics, and Traces for Unified Observability
 
Uses of Data Lakes
Uses of Data Lakes Uses of Data Lakes
Uses of Data Lakes
 
Multi-modal sources for predictive modeling using deep learning
Multi-modal sources for predictive modeling using deep learningMulti-modal sources for predictive modeling using deep learning
Multi-modal sources for predictive modeling using deep learning
 
Using Implicit Preference Relations to Improve Content-based Recommendations,...
Using Implicit Preference Relations to Improve Content-based Recommendations,...Using Implicit Preference Relations to Improve Content-based Recommendations,...
Using Implicit Preference Relations to Improve Content-based Recommendations,...
 
Feature drift monitoring as a service for machine learning models at scale
Feature drift monitoring as a service for machine learning models at scaleFeature drift monitoring as a service for machine learning models at scale
Feature drift monitoring as a service for machine learning models at scale
 

Recently uploaded

Call Girls In Mahipalpur O9654467111 Escorts Service
Call Girls In Mahipalpur O9654467111  Escorts ServiceCall Girls In Mahipalpur O9654467111  Escorts Service
Call Girls In Mahipalpur O9654467111 Escorts ServiceSapana Sha
 
Midocean dropshipping via API with DroFx
Midocean dropshipping via API with DroFxMidocean dropshipping via API with DroFx
Midocean dropshipping via API with DroFxolyaivanovalion
 
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Callshivangimorya083
 
CebaBaby dropshipping via API with DroFX.pptx
CebaBaby dropshipping via API with DroFX.pptxCebaBaby dropshipping via API with DroFX.pptx
CebaBaby dropshipping via API with DroFX.pptxolyaivanovalion
 
Ravak dropshipping via API with DroFx.pptx
Ravak dropshipping via API with DroFx.pptxRavak dropshipping via API with DroFx.pptx
Ravak dropshipping via API with DroFx.pptxolyaivanovalion
 
100-Concepts-of-AI by Anupama Kate .pptx
100-Concepts-of-AI by Anupama Kate .pptx100-Concepts-of-AI by Anupama Kate .pptx
100-Concepts-of-AI by Anupama Kate .pptxAnupama Kate
 
EMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM TRACKING WITH GOOGLE ANALYTICS.pptx
EMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM  TRACKING WITH GOOGLE ANALYTICS.pptxEMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM  TRACKING WITH GOOGLE ANALYTICS.pptx
EMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM TRACKING WITH GOOGLE ANALYTICS.pptxthyngster
 
Smarteg dropshipping via API with DroFx.pptx
Smarteg dropshipping via API with DroFx.pptxSmarteg dropshipping via API with DroFx.pptx
Smarteg dropshipping via API with DroFx.pptxolyaivanovalion
 
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...Suhani Kapoor
 
Invezz.com - Grow your wealth with trading signals
Invezz.com - Grow your wealth with trading signalsInvezz.com - Grow your wealth with trading signals
Invezz.com - Grow your wealth with trading signalsInvezz1
 
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130Suhani Kapoor
 
BigBuy dropshipping via API with DroFx.pptx
BigBuy dropshipping via API with DroFx.pptxBigBuy dropshipping via API with DroFx.pptx
BigBuy dropshipping via API with DroFx.pptxolyaivanovalion
 
Industrialised data - the key to AI success.pdf
Industrialised data - the key to AI success.pdfIndustrialised data - the key to AI success.pdf
Industrialised data - the key to AI success.pdfLars Albertsson
 
Delhi Call Girls CP 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
Delhi Call Girls CP 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip CallDelhi Call Girls CP 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
Delhi Call Girls CP 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Callshivangimorya083
 
VIP Call Girls in Amravati Aarohi 8250192130 Independent Escort Service Amravati
VIP Call Girls in Amravati Aarohi 8250192130 Independent Escort Service AmravatiVIP Call Girls in Amravati Aarohi 8250192130 Independent Escort Service Amravati
VIP Call Girls in Amravati Aarohi 8250192130 Independent Escort Service AmravatiSuhani Kapoor
 
Brighton SEO | April 2024 | Data Storytelling
Brighton SEO | April 2024 | Data StorytellingBrighton SEO | April 2024 | Data Storytelling
Brighton SEO | April 2024 | Data StorytellingNeil Barnes
 
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...Suhani Kapoor
 
Mature dropshipping via API with DroFx.pptx
Mature dropshipping via API with DroFx.pptxMature dropshipping via API with DroFx.pptx
Mature dropshipping via API with DroFx.pptxolyaivanovalion
 
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdfMarket Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdfRachmat Ramadhan H
 

Recently uploaded (20)

Call Girls In Mahipalpur O9654467111 Escorts Service
Call Girls In Mahipalpur O9654467111  Escorts ServiceCall Girls In Mahipalpur O9654467111  Escorts Service
Call Girls In Mahipalpur O9654467111 Escorts Service
 
Midocean dropshipping via API with DroFx
Midocean dropshipping via API with DroFxMidocean dropshipping via API with DroFx
Midocean dropshipping via API with DroFx
 
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
꧁❤ Greater Noida Call Girls Delhi ❤꧂ 9711199171 ☎️ Hard And Sexy Vip Call
 
CebaBaby dropshipping via API with DroFX.pptx
CebaBaby dropshipping via API with DroFX.pptxCebaBaby dropshipping via API with DroFX.pptx
CebaBaby dropshipping via API with DroFX.pptx
 
꧁❤ Aerocity Call Girls Service Aerocity Delhi ❤꧂ 9999965857 ☎️ Hard And Sexy ...
꧁❤ Aerocity Call Girls Service Aerocity Delhi ❤꧂ 9999965857 ☎️ Hard And Sexy ...꧁❤ Aerocity Call Girls Service Aerocity Delhi ❤꧂ 9999965857 ☎️ Hard And Sexy ...
꧁❤ Aerocity Call Girls Service Aerocity Delhi ❤꧂ 9999965857 ☎️ Hard And Sexy ...
 
Ravak dropshipping via API with DroFx.pptx
Ravak dropshipping via API with DroFx.pptxRavak dropshipping via API with DroFx.pptx
Ravak dropshipping via API with DroFx.pptx
 
100-Concepts-of-AI by Anupama Kate .pptx
100-Concepts-of-AI by Anupama Kate .pptx100-Concepts-of-AI by Anupama Kate .pptx
100-Concepts-of-AI by Anupama Kate .pptx
 
EMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM TRACKING WITH GOOGLE ANALYTICS.pptx
EMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM  TRACKING WITH GOOGLE ANALYTICS.pptxEMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM  TRACKING WITH GOOGLE ANALYTICS.pptx
EMERCE - 2024 - AMSTERDAM - CROSS-PLATFORM TRACKING WITH GOOGLE ANALYTICS.pptx
 
Smarteg dropshipping via API with DroFx.pptx
Smarteg dropshipping via API with DroFx.pptxSmarteg dropshipping via API with DroFx.pptx
Smarteg dropshipping via API with DroFx.pptx
 
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
VIP High Profile Call Girls Amravati Aarushi 8250192130 Independent Escort Se...
 
Invezz.com - Grow your wealth with trading signals
Invezz.com - Grow your wealth with trading signalsInvezz.com - Grow your wealth with trading signals
Invezz.com - Grow your wealth with trading signals
 
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
VIP Call Girls Service Miyapur Hyderabad Call +91-8250192130
 
BigBuy dropshipping via API with DroFx.pptx
BigBuy dropshipping via API with DroFx.pptxBigBuy dropshipping via API with DroFx.pptx
BigBuy dropshipping via API with DroFx.pptx
 
Industrialised data - the key to AI success.pdf
Industrialised data - the key to AI success.pdfIndustrialised data - the key to AI success.pdf
Industrialised data - the key to AI success.pdf
 
Delhi Call Girls CP 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
Delhi Call Girls CP 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip CallDelhi Call Girls CP 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
Delhi Call Girls CP 9711199171 ☎✔👌✔ Whatsapp Hard And Sexy Vip Call
 
VIP Call Girls in Amravati Aarohi 8250192130 Independent Escort Service Amravati
VIP Call Girls in Amravati Aarohi 8250192130 Independent Escort Service AmravatiVIP Call Girls in Amravati Aarohi 8250192130 Independent Escort Service Amravati
VIP Call Girls in Amravati Aarohi 8250192130 Independent Escort Service Amravati
 
Brighton SEO | April 2024 | Data Storytelling
Brighton SEO | April 2024 | Data StorytellingBrighton SEO | April 2024 | Data Storytelling
Brighton SEO | April 2024 | Data Storytelling
 
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
 
Mature dropshipping via API with DroFx.pptx
Mature dropshipping via API with DroFx.pptxMature dropshipping via API with DroFx.pptx
Mature dropshipping via API with DroFx.pptx
 
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdfMarket Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
Market Analysis in the 5 Largest Economic Countries in Southeast Asia.pdf
 

IRIS.TV Talks Future of Video Personalization at Cross Campus LA

  • 1. LA Machine Learning Group September 15, 2014
  • 2. About IRIS.TV Leading in-player video recommendation engine designed to create continuous streams of personalized video SaaS based solution for publishers and video technology companies looking to increase views per viewing session. IRIS.TV enables publishers to: –  connect to their audience –  increase video views –  control their programming to maximize reach and revenue potential of every viewer
  • 3. How it Works BEFORE IRIS.TV User has to navigate back to menu or app to search for more content
  • 4. How it Works Powered by Adaptive StreamTM, IRIS.TV •  Adaptive machine learning •  Programmatic video delivery •  Personalizing experiences for audiences while enforcing editorial and business rules for publishers •  Improving content discovery in a manner that facilitates audience growth and interaction •  Re-inventing the way users consume video
  • 5. How it Works Viewers of IRIS.TV enabled video players WATCH MORE, WATCH LONGER, WATCH OFTEN AFTER IRIS.TV INTEGRATION Video publishers are able to engage users for longer and serve more ads
  • 6. Speakers Tom Sullivan Chief Data Scientist Joel Spitalnik VP Engineering
  • 7. Agenda Personalization Response Curve Definition of Terms Data Science Engineering Summary
  • 8. Personalization-Value Relationship Personalization Value No Response = Unemployment Positive Correlation Something else
  • 9. Relationship Between Personalization and Engagement Stable Loss of value due to “creepiness” and loss of serendipity Strict increase in value (not likely given constraints) Value Personalization
  • 10. Definitions Asset: A good that may be represented in an online environment and consumed in computerized or physical form Examples: Video, News article, Image, Sound file, Coupon, Book Track: An ordered collection of co-presented assets Anchor Asset: The first asset presented in a track (via landing or organic search) Present Deliver Consume (in part or whole)Select
  • 11. Definitions Experience All tracks that result from interactions with a distinct anchor asset Engagement A user session that begins with the first track and ends after the final asset is presented and/or consumed Learning is based on track/experience/engagement history Asset  A   (Anchor)   Asset  D   Asset  F   Asset  A   (Anchor)   Asset  G   Asset  I   Asset  B   (Anchor)   Asset  Q   Asset  T   Dynamic Update TRACK A1 TRACK A2 TRACK B1 EXPERIENCEEXPERIENCE ENGAGEMENT
  • 12. Hybrid (front/back end) Personalization IRIS.TV ApproachValue Personalization  P0                                P1                                              P2                                        P3              P4                  P5  
  • 13. The Constrained Optimization Problem Provide personalized recommendations, P, that maximize the value of variable on the vertical axis of the response curve, V (value) Subject to constraints (including estimation risk), : Λ While remaining aware of residual benefits, R ●  The data scientist's desire execute machine learning to improve accuracy of predictions (recommended assets) ●  The opportunity to generate residual value based business insights That can be computed given the available resource environment and acceptable latency in updates to tracks, E max (V + R) = f(P) s.t. Λ,Ε “A sensible estimate is an interval estimate” IRIS.TV uses the “Strategy first” approach rather than a “data first” approach
  • 14. Building Personalized Tracks For any given anchor asset, a pre-calculated track is based on an evaluation of pairwise asset similarities reflecting a linear combination of An asset-to-asset structural component: a weighted composite similarities among the k asset features s(1) A user behavior component based on historical interactions with presented tracks s(2) Dynamically adjusted in based on user interactions with current track
  • 15. Track Construction is a Clustering Problem Note: similarities may not necessarily be symmetric when using user behavior data Asset j (in track) Present Consume Like: Dislike Skips Fully Consume Similarity S(2) 2 10,345 95.1% 7.2 : 1 11.6% 77.3% S(2) 1,2 7 4,235 9.1% 1 : 27.9 54.3% 15.9% S(2) 1,7 Sample of Historical Data Related to Paired Asset Engagements for Anchor Asset 1
  • 16. Temporal Observation Windows Temporal user behavior data may perish over time In our machine learning “sandbox”, there is little need to store and process data that has “aged” beyond some threshold The “optimal” temporal window may be inferred by using subject matter expertise and/or by ML We only consider the sufficient statistics, aggregated from the full log data and updated. This reduces the amount of data in the current environment and may improve computational efficiency, E max (V + R) = f(P) s.t. Λ, Ε Age   Obs  weight   t*   1   0  
  • 17. Preparing the data for similarity calculations API Aggregation Queries Summary observation data in temporal window, constraints, asset metadata Raw Log Data
  • 18. The Data Scientist at IRIS.TV Skills and Experience ●  Ability to program in R ●  Understanding of MySQL ●  Some familiarity with working with large, diverse data sets ●  Can move data between Hive/MySQL/R Differentiator ●  Primary Strength is Statistics v. Data Management ●  Understands dimension reduction and the pros/cons of using various methods ●  Choices are driven by objective function max (V + R) = f(P) s.t. Λ,Ε Join us?: www.iris.tv Email: tom@iris.tv joel@iris.tv
  • 19. Feature 1: Keyword Overlap Feature 2: Asset Length Tracks are Sensitive to Choice of Methods Asset A (anchor) Asset B Asset C A B C D E Asset A (anchor) Asset B Asset E Asset A (anchor) Asset B Asset D Single-link 2NN Complete Complete method: •  leads to possibly fewer computations (when n is sufficiently large) •  Can result more spherical clusters with respect to anchor asset •  Increases “track strength”
  • 20. Increasing Personalization: Asset-to-Asset Feature Weighting •  Which features are more relevant in the user’s decision to consume assets? •  Pairwise similarity, between two assets s(1) i,j are a weighted sum of k separate feature- level similarities from asset metadata (e.g. length, keyword overlap, genre, publication date, mood, etc.) •  We use simulated annealing, simplex-marching, tree-based starting points, parsimony penalty, and loss values driven by (E,R)
  • 21. Increasing Personalization with Groups The user population may be partitioned into groups based on observable characteristics and revealed feature preferences, each having a distinct set of composite similarities – leading to possibility of different tracks for each group Personalization-driven user partitioning based on location (and possibly device type) to improve “cold start” Other  Fans   World Asia  Europe   S.  America   Not  Brazil   N.  America   USA   Not  USA   Red  Sox  Fans   Yankee  Fans  Bots   Brazil  
  • 22. Bots and Bias •  Not all user engagement data is useful – there are non-human users whose influence we try to remove / downweight the user behavior component of similarity computation •  e.g. One user selected 1600 videos in 10 minutes •  In polarizing topics such as sports, politics, etc., and quality assessments there may be bias in the like /dislike actions Braz   USA   Red  Sox  Fans   Yankee  Fans   Other  Fans  Bots   X  
  • 23. Bucket Re-Evaluation Bucket assignment is a fuzzy classification problem and we assign a user to the bucket with the highest membership probability Periodically, we revisit the buckets and memberships •  If every bucket has only one user, the backend has generated personalization in its highest form (with respect to the IRIS.TV definition) •  As more data becomes available on a user based on their historical engagements, their bucket assignment may change
  • 24. Hybrid (front/back-end) Personalization, so far Value Personalization  P0                                    P1                                                              P2                                              P3                    P4                        P5   Anchoring +similarity   +  Historical  user   data   +  buckets   +beVer  buckets   +real-­‐Xme   feedback   Next  up:  Joel  Spitalnik,  to  talk   about  the  technology  stack,  NLP,   and  dynamic  track  adjustment    
  • 25. Front End Origins of Recommendation System JukeboxTV Manual Curation Building Data Structures Data Munging Natural Language Processing Mapping Keywords Service Oriented Architecture Agile Development Minimizing Latency
  • 26. Preparing the data for similarity calculations Online Rec System Aggregation Queries Summary observation data in temporal window, constraints, asset metadata Raw Log Data API
  • 27. Thank You Questions? www.iris.tv tom@iris.tv joel@iris.tv rohan@iris.tv