SlideShare a Scribd company logo
1 of 33
Download to read offline
Nir Yungster
Kamil Sindi
Building turn-key
recommendations for 5% of
internet video
1 About JW Player
Building a Recommender Service
Improving the Recommender
Future Directions
2
3
4
Agenda
About JW Player
Part 1
About JW Player
● Open source video player + video
platform
● 5% of all video plays on the web
● Per month:
○ 40Bn plays
○ 100 TB events
● 15K Customers
PLAYER Data
Analytics
The fastest online
video player
(2008)
Data-driven products (e.g.
Recommendations)
(2016)
Dashboards, Audience
Measurement
(2014)
Data Has Become Core to JW Player’s Strategy
Video Management
and Delivery
(2011)
PLATFORM
Building a
Recommender
Service
Part 2
Increases views, engagement and ad revenue with minimal effort or
investment by publisher
JW Recommendations
MVP Focused on Product Reqs and Scalability
● 20K requests per second
● Support legacy endpoints
○ Non-recommendations playlists
● Business rule features (e.g. sunrise, sunset, geo block)
● Include video metadata in response (conversions, manifest, etc.)
● Pass product “sniff test”
● Rudimentary A/B testing using click-through rates
○ Beat random
Data Types At Our Disposal for Recommending
Association-based
recommendations
Content-based
recommendations
(& Trending videos)
Title: Top ten Snowboarding
Destinations in Colorado
Description, keywords
● Association → Association Rule Mining
○ Viewers who watched X also watched Y
We Layered Classic Algorithms That Were Easy to
Implement
● Content → BM25 (think tf-idf)
○ Elasticsearch
● Trending
○ Exp. weighted moving avg of plays
Rec 1: “Best hotels in Boulder”
Rec 2: “Amazing 1080”
Rec 3: “Best ski slopes in Colorado”
Rec 4: “Snowboarding is fun!”
Rec 5: “Top Snowboarding schools”
Rec 6: “Kardashian Katastrophe!”
Rec 7: “Cats on Skis”
Top ten Snowboarding
Destinations in Colorado, 2018
Example Recommendations
Similar titles
Highly
co-watched
Trending
Association Pipeline Content Pipeline
Architecture
Results: We Met Goals :-)
✓ 20K requests per second
✓ Support legacy endpoints
✓ Business rule features (e.g. sunset, sunrise, geo block)
✓ Include video metadata in response (conversions, manifest, etc.)
○ Use log-based architecture to sync from various sources
✓ Pass product “sniff test”
✓ Rudimentary A/B testing
○ Beat random when looking at Overlay Click-Through Rate
○ Bested competitors in customer-led A/B tests
Beyond the MVP
How can we drive more value to customers?
How can we continue to grow competitive advantage?
Improving the
Recommender
Part 3
Wait, What Exactly Are We Improving?
Click Through Rate Completion Rate
Ad Impressions Viewer Time
Americans spend 2+ hrs on social media
Viewer Time, the Unit of Online Currency
Our publishers are fighting for time
Recommendations can drive viewer
time by either:
● More Time per Session
● More Sessions (higher retention)
First, We Need Ability to Run Experiments
● Keep viewers in consistent
variant to measure:
○ Time/session
○ Viewer retention
A/B results (JW model vs random)
● 50% more time per session on recommended content
● 10% higher viewer retention (D1, D7)
The Natural Itch to Test Stuff
We can now run experiments and understanding
impact on viewer time
Hypothesis
“If we boost recently
produced content,
recs will be more
relevant”
Experiment
What happens to
time spent?
Some of the Initial Tests That Were Tried
Experiment Result
Recommendation Algorithm (hypothesis)
Swap in Word2Vec title similarity instead of tf-idf
Boost recent content
Try trending only
Try different ordering of layers
2 Weeks
3 Weeks
1 Week
2 Weeks
Offline Testing = Faster Iteration
Fast Iteration Cycles
Build
Signals
Training
Data
Model
Model
Output
Predict
Evaluation,
validation
Improve Features,
Model, Data
Run Experiment
Build
Recommendation Algorithm (hypothesis)
Choosing An Offline Performance Metric
● Time spent in a session aggregates behavior over a sequence of
recommendations
○ Predicting that directly is hard
● Pick closely related metric to measure effectiveness of a single
recommendation
○ Time watched, percent watched?
○ Probability of an “engaged watch”
Video 1 Video 2
Pairwise Empirical Engagement Rate
(PEER Score)
PEER Score = Wilson Score ( )
% video 2 watches >= 30 seconds
Metric for List of Recommended Videos V :
nDCG (V), where PEER is relevance metric
● Significant improvement
to time watched
○ 10% - 40% increase
● Improved CTR too
Success!
A/B Testing Learnings: Publishers Matter
A/B Testing Learnings: Publishers Matter
● Algorithm performance
○ Association vs Content
○ Optimal Training Window
● Publishers with viral events that affect results
○ Test results change with such events
● Publisher quirks
○ Player, Recommendations implementation
Future Directions
Part 4
...it involves deep learning
● Algorithmic Perspective
○ More Context
○ Personalization
○ Progress in deep learning for recs
● Implementation / Maintainability
○ Single Unified Model (for widely varying publishers)
○ Flexible inputs (Anything2Vec)
Deep Learning Makes Sense For Us
We’ve Taken Some Good Initial Steps
● Built and A/B tested Tensorflow
model that performs on par with
our current algorithms
● Same context, unpersonalized
● AWS SageMaker used for training
on GPUs, serving model via
Tensorflow Serving
● Trained using triplet loss to learn
video embeddings
Anchor
Positive
Example
Negative
Example
FaceNet: A Unified Embedding for Face Recognition and
Clustering (2015)
Next Challenges
● Modeling
○ Score individual videos vs. learn to rank
○ How to choose positive & negative training samples?
○ Relevance metric for hyperparameter tuning
● Architecture
○ API traffic
○ Viewer profile service
○ Tensorflow is free, but scaling it is not
Takeaways
● “Just build” can work great for MVP recommender
● Offline testing critical for algorithmic improvement
● Finding the right offline metric is key
Acknowledgements
Data Science
Graham Edge
Matthew Yu
Rik Heijdens
Bobby Han
Engineering
Doug Shore
Alex Halter
Linda Cai
Dan Meng
Leo Yu
Franklin Dement
Thank you!

More Related Content

Similar to Building turn-key recommendations for 5% of internet video

How to Master UserZoom’s Latest Study Builder & Extract Meaningful UX Insights
How to Master UserZoom’s Latest Study Builder & Extract Meaningful UX InsightsHow to Master UserZoom’s Latest Study Builder & Extract Meaningful UX Insights
How to Master UserZoom’s Latest Study Builder & Extract Meaningful UX Insights
UserZoom
 
The Art of the Minimum Viable Product (MVP)
The Art of the Minimum Viable Product (MVP)The Art of the Minimum Viable Product (MVP)
The Art of the Minimum Viable Product (MVP)
Movel
 

Similar to Building turn-key recommendations for 5% of internet video (20)

[Webinar] Getting started with server-side testing - presented by WiderFunnel...
[Webinar] Getting started with server-side testing - presented by WiderFunnel...[Webinar] Getting started with server-side testing - presented by WiderFunnel...
[Webinar] Getting started with server-side testing - presented by WiderFunnel...
 
User testing methodology
User testing methodologyUser testing methodology
User testing methodology
 
UserZoom Education Series - Research Deep Dive - Advanced - Task-Based TOL (b...
UserZoom Education Series - Research Deep Dive - Advanced - Task-Based TOL (b...UserZoom Education Series - Research Deep Dive - Advanced - Task-Based TOL (b...
UserZoom Education Series - Research Deep Dive - Advanced - Task-Based TOL (b...
 
Userlytics User Testing: The Basics
Userlytics User Testing: The BasicsUserlytics User Testing: The Basics
Userlytics User Testing: The Basics
 
Software Release Orchestration and the Enterprise
Software Release Orchestration and the EnterpriseSoftware Release Orchestration and the Enterprise
Software Release Orchestration and the Enterprise
 
Better Together: Player + Analytics Webinar
Better Together: Player + Analytics WebinarBetter Together: Player + Analytics Webinar
Better Together: Player + Analytics Webinar
 
A/B Testing for WordPress & Drupal
A/B Testing for WordPress & DrupalA/B Testing for WordPress & Drupal
A/B Testing for WordPress & Drupal
 
YouTube SEO ( A guide to YouTube SEO)
YouTube SEO ( A guide to YouTube SEO)YouTube SEO ( A guide to YouTube SEO)
YouTube SEO ( A guide to YouTube SEO)
 
How to Effectively Experiment in PM by LendingTree Sr PM
How to Effectively Experiment in PM by LendingTree Sr PMHow to Effectively Experiment in PM by LendingTree Sr PM
How to Effectively Experiment in PM by LendingTree Sr PM
 
Optimal Search Engine Marketing
Optimal Search Engine MarketingOptimal Search Engine Marketing
Optimal Search Engine Marketing
 
Final 2013-su5-knewledge-you tube
Final 2013-su5-knewledge-you tubeFinal 2013-su5-knewledge-you tube
Final 2013-su5-knewledge-you tube
 
Startupfest 2015: KRISHNA GADE (Pinterest) - "How To" Stage
Startupfest 2015: KRISHNA GADE (Pinterest) - "How To" StageStartupfest 2015: KRISHNA GADE (Pinterest) - "How To" Stage
Startupfest 2015: KRISHNA GADE (Pinterest) - "How To" Stage
 
The ROI of Beta Testing
The ROI of Beta TestingThe ROI of Beta Testing
The ROI of Beta Testing
 
Successful Testing with a Lean Team
Successful Testing with a Lean TeamSuccessful Testing with a Lean Team
Successful Testing with a Lean Team
 
Product-led growth
Product-led growthProduct-led growth
Product-led growth
 
How to Master UserZoom’s Latest Study Builder & Extract Meaningful UX Insights
How to Master UserZoom’s Latest Study Builder & Extract Meaningful UX InsightsHow to Master UserZoom’s Latest Study Builder & Extract Meaningful UX Insights
How to Master UserZoom’s Latest Study Builder & Extract Meaningful UX Insights
 
The Art of the Minimum Viable Product (MVP)
The Art of the Minimum Viable Product (MVP)The Art of the Minimum Viable Product (MVP)
The Art of the Minimum Viable Product (MVP)
 
Google Analytics Video Event Tracking
Google Analytics Video Event TrackingGoogle Analytics Video Event Tracking
Google Analytics Video Event Tracking
 
Closing the Loop: Enhancing User Experience with Monetization | Tal Shoham
Closing the Loop: Enhancing User Experience with Monetization | Tal ShohamClosing the Loop: Enhancing User Experience with Monetization | Tal Shoham
Closing the Loop: Enhancing User Experience with Monetization | Tal Shoham
 
youtube.docx
youtube.docxyoutube.docx
youtube.docx
 

Recently uploaded

一比一原版布兰迪斯大学毕业证如何办理
一比一原版布兰迪斯大学毕业证如何办理一比一原版布兰迪斯大学毕业证如何办理
一比一原版布兰迪斯大学毕业证如何办理
A
 
一比一定制(USC毕业证书)美国南加州大学毕业证学位证书
一比一定制(USC毕业证书)美国南加州大学毕业证学位证书一比一定制(USC毕业证书)美国南加州大学毕业证学位证书
一比一定制(USC毕业证书)美国南加州大学毕业证学位证书
Fir
 
一比一原版(毕业证书)新西兰怀特克利夫艺术设计学院毕业证原件一模一样
一比一原版(毕业证书)新西兰怀特克利夫艺术设计学院毕业证原件一模一样一比一原版(毕业证书)新西兰怀特克利夫艺术设计学院毕业证原件一模一样
一比一原版(毕业证书)新西兰怀特克利夫艺术设计学院毕业证原件一模一样
AS
 
一比一原版(UWE毕业证书)西英格兰大学毕业证原件一模一样
一比一原版(UWE毕业证书)西英格兰大学毕业证原件一模一样一比一原版(UWE毕业证书)西英格兰大学毕业证原件一模一样
一比一原版(UWE毕业证书)西英格兰大学毕业证原件一模一样
Fi
 
原版定制美国加州大学河滨分校毕业证原件一模一样
原版定制美国加州大学河滨分校毕业证原件一模一样原版定制美国加州大学河滨分校毕业证原件一模一样
原版定制美国加州大学河滨分校毕业证原件一模一样
A
 
一比一原版(Soton毕业证书)南安普顿大学毕业证原件一模一样
一比一原版(Soton毕业证书)南安普顿大学毕业证原件一模一样一比一原版(Soton毕业证书)南安普顿大学毕业证原件一模一样
一比一原版(Soton毕业证书)南安普顿大学毕业证原件一模一样
Fi
 
Production 2024 sunderland culture final - Copy.pptx
Production 2024 sunderland culture final - Copy.pptxProduction 2024 sunderland culture final - Copy.pptx
Production 2024 sunderland culture final - Copy.pptx
ChloeMeadows1
 
原版定制(爱大毕业证书)英国爱丁堡大学毕业证原件一模一样
原版定制(爱大毕业证书)英国爱丁堡大学毕业证原件一模一样原版定制(爱大毕业证书)英国爱丁堡大学毕业证原件一模一样
原版定制(爱大毕业证书)英国爱丁堡大学毕业证原件一模一样
gfhdsfr
 
一比一原版(TRU毕业证书)温哥华社区学院毕业证如何办理
一比一原版(TRU毕业证书)温哥华社区学院毕业证如何办理一比一原版(TRU毕业证书)温哥华社区学院毕业证如何办理
一比一原版(TRU毕业证书)温哥华社区学院毕业证如何办理
Fir
 
一比一原版(Polytechnic毕业证书)新加坡理工学院毕业证原件一模一样
一比一原版(Polytechnic毕业证书)新加坡理工学院毕业证原件一模一样一比一原版(Polytechnic毕业证书)新加坡理工学院毕业证原件一模一样
一比一原版(Polytechnic毕业证书)新加坡理工学院毕业证原件一模一样
AS
 
一比一定制加州大学欧文分校毕业证学位证书
一比一定制加州大学欧文分校毕业证学位证书一比一定制加州大学欧文分校毕业证学位证书
一比一定制加州大学欧文分校毕业证学位证书
A
 
原版定制英国赫瑞瓦特大学毕业证原件一模一样
原版定制英国赫瑞瓦特大学毕业证原件一模一样原版定制英国赫瑞瓦特大学毕业证原件一模一样
原版定制英国赫瑞瓦特大学毕业证原件一模一样
AS
 

Recently uploaded (20)

一比一原版布兰迪斯大学毕业证如何办理
一比一原版布兰迪斯大学毕业证如何办理一比一原版布兰迪斯大学毕业证如何办理
一比一原版布兰迪斯大学毕业证如何办理
 
一比一定制(USC毕业证书)美国南加州大学毕业证学位证书
一比一定制(USC毕业证书)美国南加州大学毕业证学位证书一比一定制(USC毕业证书)美国南加州大学毕业证学位证书
一比一定制(USC毕业证书)美国南加州大学毕业证学位证书
 
一比一原版(毕业证书)新西兰怀特克利夫艺术设计学院毕业证原件一模一样
一比一原版(毕业证书)新西兰怀特克利夫艺术设计学院毕业证原件一模一样一比一原版(毕业证书)新西兰怀特克利夫艺术设计学院毕业证原件一模一样
一比一原版(毕业证书)新西兰怀特克利夫艺术设计学院毕业证原件一模一样
 
TORTOGEL TELAH MENJADI SALAH SATU PLATFORM PERMAINAN PALING FAVORIT.
TORTOGEL TELAH MENJADI SALAH SATU PLATFORM PERMAINAN PALING FAVORIT.TORTOGEL TELAH MENJADI SALAH SATU PLATFORM PERMAINAN PALING FAVORIT.
TORTOGEL TELAH MENJADI SALAH SATU PLATFORM PERMAINAN PALING FAVORIT.
 
一比一原版(UWE毕业证书)西英格兰大学毕业证原件一模一样
一比一原版(UWE毕业证书)西英格兰大学毕业证原件一模一样一比一原版(UWE毕业证书)西英格兰大学毕业证原件一模一样
一比一原版(UWE毕业证书)西英格兰大学毕业证原件一模一样
 
Thank You Luv I’ll Never Walk Alone Again T shirts
Thank You Luv I’ll Never Walk Alone Again T shirtsThank You Luv I’ll Never Walk Alone Again T shirts
Thank You Luv I’ll Never Walk Alone Again T shirts
 
GOOGLE Io 2024 At takes center stage.pdf
GOOGLE Io 2024 At takes center stage.pdfGOOGLE Io 2024 At takes center stage.pdf
GOOGLE Io 2024 At takes center stage.pdf
 
原版定制美国加州大学河滨分校毕业证原件一模一样
原版定制美国加州大学河滨分校毕业证原件一模一样原版定制美国加州大学河滨分校毕业证原件一模一样
原版定制美国加州大学河滨分校毕业证原件一模一样
 
Registry Data Accuracy Improvements, presented by Chimi Dorji at SANOG 41 / I...
Registry Data Accuracy Improvements, presented by Chimi Dorji at SANOG 41 / I...Registry Data Accuracy Improvements, presented by Chimi Dorji at SANOG 41 / I...
Registry Data Accuracy Improvements, presented by Chimi Dorji at SANOG 41 / I...
 
AI Generated 3D Models | AI 3D Model Generator
AI Generated 3D Models | AI 3D Model GeneratorAI Generated 3D Models | AI 3D Model Generator
AI Generated 3D Models | AI 3D Model Generator
 
🍑👄Dehradun Esℂorts Serviℂe☎️9315791090🍑👄 ℂall Girl serviℂe in ☎️Dehradun ℂall...
🍑👄Dehradun Esℂorts Serviℂe☎️9315791090🍑👄 ℂall Girl serviℂe in ☎️Dehradun ℂall...🍑👄Dehradun Esℂorts Serviℂe☎️9315791090🍑👄 ℂall Girl serviℂe in ☎️Dehradun ℂall...
🍑👄Dehradun Esℂorts Serviℂe☎️9315791090🍑👄 ℂall Girl serviℂe in ☎️Dehradun ℂall...
 
The Rise of Subscription-Based Digital Services.pdf
The Rise of Subscription-Based Digital Services.pdfThe Rise of Subscription-Based Digital Services.pdf
The Rise of Subscription-Based Digital Services.pdf
 
一比一原版(Soton毕业证书)南安普顿大学毕业证原件一模一样
一比一原版(Soton毕业证书)南安普顿大学毕业证原件一模一样一比一原版(Soton毕业证书)南安普顿大学毕业证原件一模一样
一比一原版(Soton毕业证书)南安普顿大学毕业证原件一模一样
 
Production 2024 sunderland culture final - Copy.pptx
Production 2024 sunderland culture final - Copy.pptxProduction 2024 sunderland culture final - Copy.pptx
Production 2024 sunderland culture final - Copy.pptx
 
iThome_CYBERSEC2024_Drive_Into_the_DarkWeb
iThome_CYBERSEC2024_Drive_Into_the_DarkWebiThome_CYBERSEC2024_Drive_Into_the_DarkWeb
iThome_CYBERSEC2024_Drive_Into_the_DarkWeb
 
原版定制(爱大毕业证书)英国爱丁堡大学毕业证原件一模一样
原版定制(爱大毕业证书)英国爱丁堡大学毕业证原件一模一样原版定制(爱大毕业证书)英国爱丁堡大学毕业证原件一模一样
原版定制(爱大毕业证书)英国爱丁堡大学毕业证原件一模一样
 
一比一原版(TRU毕业证书)温哥华社区学院毕业证如何办理
一比一原版(TRU毕业证书)温哥华社区学院毕业证如何办理一比一原版(TRU毕业证书)温哥华社区学院毕业证如何办理
一比一原版(TRU毕业证书)温哥华社区学院毕业证如何办理
 
一比一原版(Polytechnic毕业证书)新加坡理工学院毕业证原件一模一样
一比一原版(Polytechnic毕业证书)新加坡理工学院毕业证原件一模一样一比一原版(Polytechnic毕业证书)新加坡理工学院毕业证原件一模一样
一比一原版(Polytechnic毕业证书)新加坡理工学院毕业证原件一模一样
 
一比一定制加州大学欧文分校毕业证学位证书
一比一定制加州大学欧文分校毕业证学位证书一比一定制加州大学欧文分校毕业证学位证书
一比一定制加州大学欧文分校毕业证学位证书
 
原版定制英国赫瑞瓦特大学毕业证原件一模一样
原版定制英国赫瑞瓦特大学毕业证原件一模一样原版定制英国赫瑞瓦特大学毕业证原件一模一样
原版定制英国赫瑞瓦特大学毕业证原件一模一样
 

Building turn-key recommendations for 5% of internet video

  • 1. Nir Yungster Kamil Sindi Building turn-key recommendations for 5% of internet video
  • 2. 1 About JW Player Building a Recommender Service Improving the Recommender Future Directions 2 3 4 Agenda
  • 4. About JW Player ● Open source video player + video platform ● 5% of all video plays on the web ● Per month: ○ 40Bn plays ○ 100 TB events ● 15K Customers
  • 5. PLAYER Data Analytics The fastest online video player (2008) Data-driven products (e.g. Recommendations) (2016) Dashboards, Audience Measurement (2014) Data Has Become Core to JW Player’s Strategy Video Management and Delivery (2011) PLATFORM
  • 7. Increases views, engagement and ad revenue with minimal effort or investment by publisher JW Recommendations
  • 8. MVP Focused on Product Reqs and Scalability ● 20K requests per second ● Support legacy endpoints ○ Non-recommendations playlists ● Business rule features (e.g. sunrise, sunset, geo block) ● Include video metadata in response (conversions, manifest, etc.) ● Pass product “sniff test” ● Rudimentary A/B testing using click-through rates ○ Beat random
  • 9. Data Types At Our Disposal for Recommending Association-based recommendations Content-based recommendations (& Trending videos) Title: Top ten Snowboarding Destinations in Colorado Description, keywords
  • 10. ● Association → Association Rule Mining ○ Viewers who watched X also watched Y We Layered Classic Algorithms That Were Easy to Implement ● Content → BM25 (think tf-idf) ○ Elasticsearch ● Trending ○ Exp. weighted moving avg of plays
  • 11. Rec 1: “Best hotels in Boulder” Rec 2: “Amazing 1080” Rec 3: “Best ski slopes in Colorado” Rec 4: “Snowboarding is fun!” Rec 5: “Top Snowboarding schools” Rec 6: “Kardashian Katastrophe!” Rec 7: “Cats on Skis” Top ten Snowboarding Destinations in Colorado, 2018 Example Recommendations Similar titles Highly co-watched Trending
  • 12. Association Pipeline Content Pipeline Architecture
  • 13. Results: We Met Goals :-) ✓ 20K requests per second ✓ Support legacy endpoints ✓ Business rule features (e.g. sunset, sunrise, geo block) ✓ Include video metadata in response (conversions, manifest, etc.) ○ Use log-based architecture to sync from various sources ✓ Pass product “sniff test” ✓ Rudimentary A/B testing ○ Beat random when looking at Overlay Click-Through Rate ○ Bested competitors in customer-led A/B tests
  • 14. Beyond the MVP How can we drive more value to customers? How can we continue to grow competitive advantage?
  • 16. Wait, What Exactly Are We Improving? Click Through Rate Completion Rate Ad Impressions Viewer Time
  • 17. Americans spend 2+ hrs on social media Viewer Time, the Unit of Online Currency Our publishers are fighting for time Recommendations can drive viewer time by either: ● More Time per Session ● More Sessions (higher retention)
  • 18. First, We Need Ability to Run Experiments ● Keep viewers in consistent variant to measure: ○ Time/session ○ Viewer retention A/B results (JW model vs random) ● 50% more time per session on recommended content ● 10% higher viewer retention (D1, D7)
  • 19. The Natural Itch to Test Stuff We can now run experiments and understanding impact on viewer time Hypothesis “If we boost recently produced content, recs will be more relevant” Experiment What happens to time spent?
  • 20. Some of the Initial Tests That Were Tried Experiment Result Recommendation Algorithm (hypothesis) Swap in Word2Vec title similarity instead of tf-idf Boost recent content Try trending only Try different ordering of layers 2 Weeks 3 Weeks 1 Week 2 Weeks
  • 21. Offline Testing = Faster Iteration Fast Iteration Cycles Build Signals Training Data Model Model Output Predict Evaluation, validation Improve Features, Model, Data Run Experiment Build Recommendation Algorithm (hypothesis)
  • 22. Choosing An Offline Performance Metric ● Time spent in a session aggregates behavior over a sequence of recommendations ○ Predicting that directly is hard ● Pick closely related metric to measure effectiveness of a single recommendation ○ Time watched, percent watched? ○ Probability of an “engaged watch”
  • 23. Video 1 Video 2 Pairwise Empirical Engagement Rate (PEER Score) PEER Score = Wilson Score ( ) % video 2 watches >= 30 seconds Metric for List of Recommended Videos V : nDCG (V), where PEER is relevance metric
  • 24. ● Significant improvement to time watched ○ 10% - 40% increase ● Improved CTR too Success!
  • 25. A/B Testing Learnings: Publishers Matter
  • 26. A/B Testing Learnings: Publishers Matter ● Algorithm performance ○ Association vs Content ○ Optimal Training Window ● Publishers with viral events that affect results ○ Test results change with such events ● Publisher quirks ○ Player, Recommendations implementation
  • 27. Future Directions Part 4 ...it involves deep learning
  • 28. ● Algorithmic Perspective ○ More Context ○ Personalization ○ Progress in deep learning for recs ● Implementation / Maintainability ○ Single Unified Model (for widely varying publishers) ○ Flexible inputs (Anything2Vec) Deep Learning Makes Sense For Us
  • 29. We’ve Taken Some Good Initial Steps ● Built and A/B tested Tensorflow model that performs on par with our current algorithms ● Same context, unpersonalized ● AWS SageMaker used for training on GPUs, serving model via Tensorflow Serving ● Trained using triplet loss to learn video embeddings Anchor Positive Example Negative Example FaceNet: A Unified Embedding for Face Recognition and Clustering (2015)
  • 30. Next Challenges ● Modeling ○ Score individual videos vs. learn to rank ○ How to choose positive & negative training samples? ○ Relevance metric for hyperparameter tuning ● Architecture ○ API traffic ○ Viewer profile service ○ Tensorflow is free, but scaling it is not
  • 31. Takeaways ● “Just build” can work great for MVP recommender ● Offline testing critical for algorithmic improvement ● Finding the right offline metric is key
  • 32. Acknowledgements Data Science Graham Edge Matthew Yu Rik Heijdens Bobby Han Engineering Doug Shore Alex Halter Linda Cai Dan Meng Leo Yu Franklin Dement