Kuaishou’s ConAlign selectively aligns biased and unbiased recommendation towers

_reachsumit · x · 2026-07-28

Kuaishou's ConAlign proposes a conditional alignment framework for recommendation debiasing. The paper argues that industrial recommenders trained on observational data suffer from bias-induced filter bubbles, but existing debiasing methods are too costly or sacrifice factual ranking performance.

ConAlign uses a discrete gating-based mechanism to selectively transfer knowledge from a biased tower to an unbiased tower. Instead of universally correcting every signal, it balances factual accuracy with unbiased preference estimation and supports real-time streaming adaptation. The authors say it is the first streaming debiasing framework deployed in a large-scale industrial recommender system using a small fraction of unbiased random traffic, with offline experiments on three real-world datasets supporting the approach.

Original post →

More from Research

Research channel →