Improving YouTube Recommender systems for big screen users
summary
YouTube runs a multi-stage recommender that serves billions and is shifting from 2015 linear models through multitask DNNs to large contextual models in 2023 that reduce outliers and retrain offline faster. Living-room signals show longer sessions, preference for long-form and autoplay, and family-friendly co-viewing on Google TV, so engineering work focuses on shared-device co-viewing and remote-driven navigation constraints. Decision rule: if TVs provide novel data build a TV-only model decision, otherwise pipe data into existing rankers; initial A/Bs determine whether a new model ships.
AI-written summary · about
chapters (11)
- 0:00Big screen recommendations intro
- 1:13YouTube’s big screen system
- 3:42YouTube on Google TV demo
- 5:08Living room user differences
- 8:23Evolution of recommendation models
- 9:46Data optimizations for TV users
- 12:21Big screen features and challenges
- 15:35ML and system improvements
- 18:06Model improvement decisions
- 19:00Future directions for big screen
- 19:53Q&A
related talks
23mML-Powered Search & Recommendation 101
42mBuilding a Unified Viewing Experience with Multi-Provider TV Recommendation
52mReinforcement Learning in Recommender Systems: Some Challenges
18mBuilding an Incrementally Trained, Local Taste Aware, Global Deep Learned Recommender System Model
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
29mBuilding Generative Content Understanding Systems
43mRecommendation Systems in Market Research
38mHow Ranker Turned Pop Culture Lists Into Personalized TV Recommendations