Serving Billions of Personalized News Feeds with AI
summary
System uses per-user probabilities for actions (click, like, share, hide) learned from historical user feedback signals and multiplies them by product-defined weights to produce a single ranking score (example: 11% click, 2.2% like, 1.1% hide → score 277). Features and labels stream into a feature joiner, models train offline and push weights to an online aggregator. Model trains on billions of examples and the pipeline serves personalized rankings in real time.
AI-written summary · about
chapters (4)
related talks
22mGenerating a Billion Personal News Feeds
32mFeed generation at ShareChat
20mPredicting content virality to drive social traffic
28mThe Power of Personalized Customer Experiences Through AI Innovations
39mMachine Learning for Relevance and Serendipity
23mML-Powered Search & Recommendation 101
35mEnhancing User Engagement Through Personalized Recommendation at Dow Jones
21mRise of the Machines: Removing the Human-in-the-Loop