Recommendations for Building Machine Learning Software

on YouTube ↗41m

" summary

Netflix runs recommendations for 80 million members in 190+ countries and reports 80% of what people watch comes from recommendations. The system uses AWS, Java, Spark/Hadoop for offline training, a offline/nearline/online split for batch, fresh and real-time updates, and EVCache/memcached for serving; models include matrix factorization, deep nets and Hogwild SGD and get validated by A/B tests measuring retention and streaming hours. Engineering rules include maximize single machine before distributing.

AI-written summary · about

" chapters (12)

1. 0:00Introduction to ML recommendations
2. 1:16Netflix recommendations evolution
3. 4:48ML software building blocks
4. 10:22System architecture design
5. 13:14Experiment in production gap
6. 16:44Modular algorithm design
7. 20:16Model input output mapping
8. 22:44Single machine optimization
9. 25:55Beyond metric testing
10. 28:42ML as software development
11. 30:05Key takeaways summary
12. 30:28Q&A session

" related talks

• 22mJustin Basilico, Research/ Engineering Manager, NetflixJustin Basilico · MLconf SF 2015
• 03mMachine Learning for PersonalizationTony Jebara · O'Reilly Artificial Intelligence Conference 2019 - New York, New York
• 18mBuilding an Incrementally Trained, Local Taste Aware, Global Deep Learned Recommender System ModelAnoop Deoras · MLconf SF 2019
• 03mInterviewVille Tuulos · Scale By The Bay 2019
• 29mXavier Amatriain, Director of Algorithms Engineering, NetflixXavier Amatriain · MLconf SF 2014
• 23mML-Powered Search & Recommendation 101Sergey Polyashov · Conf42 Machine Learning 2025
• 04mInterviewShiva Chaitanya · Scale By the Bay 2017
• 01h 31mHands on - Build a Recommender systemCamille Couturier · PyData Amsterdam 2019