PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation
Researchers have developed PEARL, a novel framework for unbiased percentile estimation in large-scale livestream recommendation systems. This method uses contrastive learning to model relative user preferences, avoiding the bias introduced by varying user activity levels. Online A/B testing on a major livestream platform showed significant improvements, including a 2.10% increase in watch duration and a 1.49% rise in interaction rates. AI
IMPACT Introduces a novel method to mitigate bias in large-scale recommendation systems, potentially improving user experience and platform engagement.