User True Interest Survey (UTIS) Model for Facebook Reels

The UTIS model augments Facebook Reels' recommendation stack by incorporating direct user feedback collected through in‑feed surveys. By converting binary interest signals into a probabilistic score, the system surfaces niche, high‑quality videos, reduces reliance on noisy engagement metrics, and drives measurable gains in user retention, satisfaction, and overall platform engagement.

Deep Technical Analysis

UTIS operates as a lightweight alignment layer that ingests the primary multi‑task ranking model’s predictions together with engineered features derived from survey responses, content metadata, and user behavior. The layer is trained on a bias‑adjusted dataset, learns to predict the probability of true interest, and outputs a scalar score that is subsequently fed back into the ranking pipeline.

Survey Data Collection and Bias Correction

Randomly selected users encounter a single‑question prompt—“How well does this video match your interests?”—rated on a 1‑5 scale. Responses are binarized (interested vs. not interested) and re‑weighted to correct for sampling and non‑response bias, producing a calibrated dataset that more accurately reflects the broader user base.

Model Architecture and Alignment Layer

The alignment layer is a shallow feed‑forward network that consumes the main model’s logits, user‑level engagement features, and content‑level attributes (audio style, production quality, mood). It is trained using a binary cross‑entropy loss to predict the binarized survey outcome. The design prioritizes interpretability; feature importance can be extracted via SHAP values to understand drivers of interest.

Integration into the Ranking Funnel

During inference, the UTIS score is combined with existing relevance signals. Videos with high predicted interest receive a modest boost, while low‑score items are demoted. Offline experiments show a 3.7 % lift in tier‑0 retention, and online A/B tests report a 2.1 % increase in average watch time per session.

Future Directions and Challenges

Ongoing work targets sparse‑user scenarios, further bias reduction, and cohort‑specific personalization. Exploration of agentic AI techniques aims to make the alignment layer more autonomous, while insights from the DeepSeek V4 large language model inform richer user representations and content semantics.