← News

ResearchSources: arXiv · 18 Nov 2026

Kuaishou-style PEARL paper describes contrastive debiasing for livestream recommendations

A new arXiv paper proposes PEARL, a method that models relative preference rather than raw engagement, reporting gains in online A/B tests on a large livestream platform.

In plain words

Recommender systems learn from user clicks and watches, but very active users can dominate that data and skew what the model learns. PEARL instead compares users against each other, estimating where a preference ranks in percentile terms rather than how large the engagement numbers are. The authors say this yields unbiased preference estimates and better recommendations.

What was announced

The paper, published on arXiv, argues that recommender systems trained on interaction data suffer from what the authors call behavioral intensity imbalance: highly active users generate disproportionate feedback signals, so observed interactions stop reflecting true preferences and model quality degrades at scale. PEARL addresses this with a nonparametric contrastive approach that approximates percentile relationships directly from real interaction samples, avoiding auxiliary distribution estimation models. The authors provide a theoretical argument that such pairwise comparisons give unbiased estimates of percentile-based preference signals.

To handle sparse and discrete feedback, the framework adds a prediction-based bootstrapping mechanism for percentile smoothing, a generalized value-weighted formulation, and a co-training strategy. In offline experiments the authors report consistent improvements across multiple ranking targets. They also describe deployment on a production livestream platform with a combined user base of billions, where online A/B testing showed a 2.10% increase in Watch Duration, 0.80% in Consumption Amount, 1.49% in Interaction Rate, and a 6.91% decrease in Report Rate. These figures come from the authors' own testing and have not been independently verified.

What it does not show

  • The source is only the paper's abstract, so the full methodology, experimental setup, and baseline comparisons cannot be assessed from this evidence.
  • The reported metrics and theoretical claims are the authors' own; no independent confirmation is available.

For security and IT teams

Status
The authors state PEARL is deployed in a production livestream platform with a combined user base of billions, with gains confirmed through online A/B testing.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.