ML-Powered Search & Recommendation 101
summary
Claims concrete numbers: recommendations drive 30% of YouTube views, 35 of Amazon purchases, and 80% of Netflix watches; production pipelines use a multi-stage funnel that narrows from 10k to 1k to 50 candidates, with lightweight GBDTs upstream and deep models later. Retrieval relies on libraries like phase and scan and HNSV plus heuristics; late-fusion item/user towers report over 110 speed up while keeping ~80% of quality. Reinforcement learning and contrastive training target long-term value.
AI-written summary · about
chapters (9)
- 0:00Introduction and Overview
- 0:55Importance of Search and Recommendation Systems
- 1:38Evaluating System Goals
- 2:32System Architecture
- 4:40Candidate Generation and Ranking
- 9:02Ranking Models and Challenges
- 15:49Design Principles for Scalability
- 20:11Trends in Recommendation Systems
- 22:40Conclusion and Final Thoughts
related talks
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
24mImproving YouTube Recommender systems for big screen users
21mRecsys Keynote: Improving Recommendation Systems & Search in the Age of LLMs
32mLet's talk about Recommendation Systems!
29mBuilding Generative Content Understanding Systems
41mRecommendations for Building Machine Learning Software
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch