Problem
Short-video recommenders are trained on likes, shares and watch time. These signals are noisy and favor content with high short-term value over content that matches what a person cares about, which costs long-term value. At Facebook Reels the existing interest heuristics identified true interests with only 48.3% precision.
Example from the article: a user is shown a Reel, and a one-question survey asks how well the video matches their interests on a 1–5 scale.
Solution in brief
Randomly sampled in-feed surveys give a direct label for interest match. A lightweight "perception layer" on top of the main ranking model predicts the probability that a user is satisfied with a video, using the main model's existing event predictions as inputs. The score feeds late-stage ranking and retrieval.
Questions this case answers
- How to train a ranking model on a sparse direct-feedback signal when the main model learns only from implicit engagement
- How to use a small randomized survey sample to correct implicit-engagement proxies for user interest in ranking and retrieval
- How to make a video recommender match users' true interests when likes and watch time are noisy proxies?
- How to use sparse in-feed survey answers to improve a large ranking model?
Tools and methods
- In-feed interest survey — one question on a random share of viewing sessions, answers weighted for sampling and nonresponse bias.
- UTIS alignment layer (User True Interest Survey) — lightweight model trained on binarized survey answers, with the main model's predictions as features.
- Late-stage ranking input — UTIS score added to the final value formula.
- Retrieval re-ranking and distillation — aggregated survey data rebuilds interest profiles; user-to-item retrieval models learn from UTIS predictions.
- A/B test — online confirmation on more than 10 million users.
Google or a one-time email code. No password.