Adapting the Facebook Reels RecSys AI Model Based on User Feedback

Meta introduces the User True Interest Survey (UTIS) model to improve Facebook Reels recommendations by incorporating direct user feedback beyond traditional engagement signals.

• •Traditional signals like watch time and likes achieve only 48.3% precision in identifying true user interests, motivating a survey-based approach
• •Daily randomized single-question surveys (1-5 scale) collect thousands of real-time interest-match responses, corrected for sampling and nonresponse bias
• •UTIS is a lightweight alignment model layer built on top of the existing large multi-task ranking model, using binarized survey responses and engineered features
• •Deployed in both Late Stage Ranking (score boost/demotion) and Early Stage Ranking (user interest profile reconstruction via knowledge distillation)
• •Online A/B tests with 10M+ users showed +5.4% high survey ratings, +5.2% total engagement, -6.84% low ratings, and -0.34% integrity violations