Machine Learning
For Recommendations
A Netflix Perspective
Faisal Siddiqi
9 Feb 2018, Make School, SF
@faisalzs
?
● > 117 million members streaming over 125
million hours a day
● > 190 countries
● > 1000 device types
● Log 100B events/day
● 35.2% of peak US downstream traffic*
* Source: https://www.sandvine.com/resources/global-internet-phenomena/2016/north-america-and-latin-america.html
Netflix Scale
Netflix’s History with ML
Recommendations (Predict Movie Ratings)
2008 - 2009
ML @ Netflix Today
Recommendations
Search
Artwork Personalization
Content Promotion
Artwork Generation
Artwork Enhancement
Asset Automation
Creative
Personalization Content
Other
Popularity
Valuation
Physical Production
Studio in the Cloud
Streaming QoE
Content Delivery
Programmatic Marketing
… and many more!
Turn on Netflix and the best content for
you starts playing automatically
Recommendations
Recommendations
● A few seconds to find something
great to watch…
● Can only show a few titles
● Enjoyment directly impacts
customer satisfaction
● Generates over $1B per year of
Netflix revenue
● How? Personalize everything
R
O
W
S
RANKING
Row
Title
Score
Hero
Image
Horizontal
Image
Everything is Personalized
Recommendations with Context
2017-12-10 15:40:22
2017-12-23 19:32:10
2017-12-24 12:05:53
2017-12-27 22:40:22
2017-12-29 19:39:36
2017-12-30 20:42:13
Context ActionSequence
per user
?
Time
Data from
Millions of
users
Training
pipelines
Models
Precompute System
Rankings
Online
caches
AB Test
Allocation
Systems & Pipelines
ML and Systems
Source: Hidden Technical Debt in Machine Learning Systems
ExplorationData Models Deploy
Training
Data
Infra
Plays
Impressions
User Actions
Store & Serve
Adhoc
Exploration
Infra
Notebooks
Libraries
Frameworks
Model
Development
Infra
Orchestration
TensorFlow
Scikit-learn
Hyper Params
Production
Deployment
Infra
Continuous Integration
Explore/Exploit
Logging Feedback
Orchestrating
ML Training
Takeaways
● Machine Learning is extensively used for Recommendations as
well as broadly across several disciplines
● Every pixel, every device, every contact presents an opportunity
for personalization
● Solid Infrastructure and Systems is a pre-req for large-scale ML
● Investing in robust software best practices is just as important for
Machine Learning as for traditional software development
Fresh.
Exciting.
Engaging.
Moments of Joy.
Conclusion
Netflix is hyper data-driven and a robust experimentation
platform facilitates fast and objective decision-making
ML innovations allow us to make the Netflix experience highly
personalized enabling member joy
Questions?
Thank you.

Machine learning for Netflix recommendations talk at SF Make School

Editor's Notes

  • #10 Collection of different algorithms that come together to create the complete Netflix experience
  • #17 Fresh, Exciting,