PEARL Framework Uses Contrastive Learning for Unbiased Recommendation
A new research paper introduces PEARL, a nonparametric contrastive percentile approximation framework designed to address behavioral intensity imbalance in recommender systems. This imbalance, caused by heterogeneous user engagement patterns, skews feedback signals and causes models to overrepresent highly active users while underrepresenting others. PEARL models relative preference signals instead of absolute engagement magnitudes, using real contrastive interaction samples to approximate percentile relationships without auxiliary distribution estimation models. The framework provides theoretical justification for its pairwise debiasing approach. The paper is available on arXiv under identifier 2605.21752.
Key facts
- PEARL addresses behavioral intensity imbalance in recommender systems.
- Behavioral intensity imbalance arises from heterogeneous user engagement patterns.
- The imbalance causes models to amplify signals from highly active users.
- PEARL uses nonparametric contrastive percentile approximation.
- It models relative preference signals rather than absolute engagement magnitudes.
- PEARL leverages real contrastive interaction samples.
- No auxiliary distribution estimation models are needed.
- The paper is on arXiv with identifier 2605.21752.
Entities
Institutions
- arXiv