Kuaishou's UAME uses uncertainty to fix satisfaction label bias in short video ranking
Uncertainty-aware model reduces label bias from noisy user signals, improving satisfaction alignment.
The paper proposes UAME, an uncertainty-aware end-to-end multi-objective ensemble ranking framework for short video recommendations. By representing predictions as Gaussian variables with quantified variance, UAME mitigates satisfaction label bias from conflicting behavioral signals like clicks and watch time. It uses a probabilistic pairwise ranking loss and an uncertainty-aware sample-level weighting scheme to mitigate bias. Deployed in a production system, UAME consistently improves the state-of-the-art paradigms EMER and EASQ and better aligns with questionnaire-based user satisfaction.
- UAME models predictions as Gaussian variables with mean (score) and variance (uncertainty) to handle noisy user signals.
- A probabilistic pairwise ranking loss and uncertainty-aware sample weighting mitigate satisfaction label bias.
- Deployed in Kuaishou's production short-video system, significantly improving alignment with user satisfaction surveys over EMER and EASQ.
Why It Matters
Better recommendation quality by explicitly modeling uncertainty in user signals, reducing bias from noisy behavioral proxies.