17/01/2026
Delivering personalized video recommendations at scale is a core challenge that impacts user satisfaction and long-term engagement. For Facebook Reels, we’ve taken a new approach to address this challenge: moving beyond engagement metrics to directly model user feedback.
Our new User True Interest Survey model leverages large-scale, randomized in-context surveys to capture real user perceptions of content relevance. By training a dedicated alignment model on this data, we’ve improved our ability to surface niche, high-quality content that matches users’ true interests.
Technical highlights:
1️⃣ Our True Interest Survey Model integrates survey-based user perception into the ranking system, improving accuracy (59.5% → 71.5%), precision (48.3% → 63.2%), and recall (45.4% → 66.1%) over heuristic baselines.
2️⃣ Large-scale A/B tests with 10M+ users show increased engagement (+5.2%), higher satisfaction (+5.4% in high survey ratings) and reduced integrity violations (-0.34%).
3️⃣ The model enables more interpretable, diverse, and personalized recommendations, even for users with sparse engagement histories.
We’re continuing to address challenges like data sparsity and bias, and are exploring advanced modeling techniques, including large language models, to further improve relevance and diversity.
Read a deep dive on our approach: https://go.meta.me/88da2f
