Facebook Engineering


Facebook Reels improved video recommendations by incorporating direct user feedback through a User True Interest Survey (UTIS) model. Traditional engagement metrics like watch time only achieved 48.3% precision in identifying true interests. The UTIS model collects real-time survey responses asking users to rate content relevance on a 1-5 scale, then uses this data to train a lightweight alignment layer integrated into both early and late-stage ranking. Results showed accuracy improved from 59.5% to 71.5%, with A/B tests demonstrating +5.4% increase in high ratings, +5.2% boost in engagement, and improved retention metrics. The approach surfaces more niche, high-quality content while reducing generic recommendations.

Adapting the Facebook Reels RecSys AI Model Based on User Feedback

Framework: User True Interest Survey (UTIS) Model

Integrating the UTIS Model in the Main Ranking System