Facebook Reels Refines Recommendations Using User Survey Feedback

TL;DR Summary
Facebook Reels enhances personalization by deploying the User True Interest Survey (UTIS), a survey-driven alignment layer atop the main ranking model. Daily in-context surveys measure how well videos match user interests, boosting niche content and engagement. Offline metrics improve from 59.5% to 71.5% in accuracy, 48.3% to 63.2% in precision, and 45.4% to 66.1% in recall; online A/B tests show higher engagement and fewer integrity issues. UTIS is integrated in late-stage ranking and early-stage retrieval via a Perception Layer to handle sparse feedback, with future work focusing on bias reduction and more advanced modeling.
Adapting the Facebook Reels RecSys AI Model Based on User Feedback Engineering at Meta
Reading Insights
Total Reads
0
Unique Readers
7
Time Saved
7 min
vs 8 min read
Condensed
94%
1,496 → 94 words
Want the full story? Read the original article
Read on Engineering at Meta