Being Smart with Art
summary
Netflix uses machine learning and A/B explore‑exploit to pick and serve artwork, claiming measurable lifts in streaming hours and valuing the recommender at over $1B. Systems run discrete and continuous bandits (UCB and Thompson sampling), log impressions versus streams to define success, and train personalized models (deep nets, logistic regression, trees) that back off to global counts for small countries. Data shows surprising winners like the Caravan image wins and large country‑by‑country aesthetic differences, and computer vision features (color, closeups, emotion) guide automated asset creation.
AI-written summary · about
chapters (31)
- 0:00Intro
- 0:38Netflix
- 1:25Recommendation System
- 2:43What are Assets
- 3:46stylized title
- 4:08billboard
- 4:40image quality
- 5:02fuller house
- 5:39first motivation
- 7:42second motivation
- 8:55review
- 9:11Back of box
- 10:10Horizontal image
- 10:45The Sixth Sense
- 11:26Before and After
- 11:48Last Tango in Halifax
- 12:02Killing Season
- 12:33Dragon Heart
- 13:06Staten Island Summer
- 14:48Explorer Exploit
- 16:52Pop Quiz
- 18:03Devices
- 18:25Downton Abbey
- 18:43Avoiding clickbait
- 19:26Diversity
- 20:02Personalized Asset Optimization
- 21:20Image Aesthetics
- 23:19Personalized Modeling
- 23:44Vision
- 24:30Emotions
- 25:44Production PostProduction
related talks
03mMachine Learning for Personalization
41mRecommendations for Building Machine Learning Software
27mWhen the Art of Entertainment ties the knot with Science
23mML-Powered Search & Recommendation 101
15mAI-Driven Application Optimization
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization
32mRebuilding the AI Platform: Designing for Today’s Chaos and Tomorrow’s Unknowns