PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation

2026-05-23 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

PEARL, a nonparametric contrastive percentile approximation framework, addresses behavioral intensity imbalance in industrial-scale livestream recommender systems. This framework, introduced in 2018, models relative preference signals directly from real contrastive interaction samples, bypassing the need for auxiliary distribution estimation models. It provides theoretical justification for unbiased percentile-based preference signals. For broader applicability, PEARL incorporates a prediction-based bootstrapping mechanism for percentile smoothing to handle sparse feedback, a generalized value-weighted formulation, and a co-training strategy to enhance modeling flexibility and representation learning. Extensive offline experiments demonstrated PEARL's effectiveness in mitigating behavioral bias and improving recommendation performance. Online A/B tests on a production livestream platform serving billions of users confirmed substantial real-world gains, including +2.10% Watch Duration, +0.80% Consumption Amount, +1.49% Interaction Rate, and -6.91% Report Rate.

Key takeaway

For Machine Learning Engineers building large-scale recommender systems, particularly for livestream platforms, you should consider PEARL to mitigate behavioral intensity bias. This framework directly models relative user preferences, improving ranking quality for all user activity levels. Implement its contrastive percentile learning, potentially with co-training, to achieve significant gains in watch duration and interaction rates, even for sparse feedback scenarios.

Key insights

PEARL uses contrastive learning to estimate unbiased percentile-based preferences, mitigating behavioral intensity bias in large-scale recommender systems.

Principles

• Relative preference signals debias heterogeneous user engagement.
• Pairwise comparisons yield unbiased percentile estimates.
• Multi-sample averaging reduces estimation variance.

Method

PEARL employs single and multi-sample contrastive learning to approximate user-specific percentiles from interaction samples. It uses prediction-based bootstrapping for sparse data and co-training for joint optimization with absolute magnitudes.

In practice

• Implement reservoir sampling for historical user data.
• Apply gradient gating for stable training with sparse users.
• Co-train percentile and regression objectives for dual benefits.

Topics

• Recommender Systems
• Contrastive Learning
• Livestream Recommendation
• Behavioral Bias
• Percentile Estimation
• MLOps

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Cracking the Cold Start Problem — Hybrid recommender systems combining collaborative filtering, embeddings, and bandit learning improve cold start and fairness.
• A Data-Augmented Contrastive Learning Approach to Nonparametric Density Estimation — DANNCE offers an efficient, consistent, and automatically normalized nonparametric density estimation via contrastive learning and data augmentation.
• Statistically Reliable LLM-Based Ranking Evaluation via Prediction-Powered Inference — PRECISE offers statistically unbiased ranking evaluation by combining limited human labels with extensive LLM judgments.
• BAHSD: Bridging the Long-tail Gap via Adaptive Distillation in Black-box Sequential Recommendation — Black-box sequential recommendation distillation benefits from adaptively handling signal heterogeneity via multi-scale probing and a hierarchical objective.
• From Candidate Retrieval to Fine-Grained Ranking (Part 1) — The article distinguishes between the two-tower model for candidate retrieval and the ranking model for fine-grained ranking within recommendation systems.
• Rethinking Semantic–Collaborative Integration in Recommenders, Efficient Token-Aware Clustering for Multivector Retrieval, and More! — Effective information retrieval systems benefit from managing complementarity, optimizing indexing, and distilling LLM utility.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.