Kuaishou proposes an uncertainty-aware ranking framework to reduce short-video label bias

_reachsumit · x · 2026-07-21

## Uncertainty-aware ranking to reduce label bias in short video systems Kuaishou presents **Uncertainty as Remedy**, a ranking framework for short-video recommendation that explicitly models conflicts among multiple proxy signals. - The goal is to reduce **satisfaction label bias** in multi-objective ensemble ranking. - The method uses uncertainty to capture disagreement across signals instead of treating all labels as equally reliable. - The authors argue this leads to better ranking decisions in production-style recommendation systems.

Original post →

More from Research

Research channel →