Finding Interest Needle in Popularity Haystack: Improving Retrieval by Modeling Item Exposure

2025-03-31

AI Summary

Recommendation systems exacerbate popularity bias through closed-loop feedback, severely limiting exposure for long-tail items. Existing approaches primarily mitigate bias during training or ranking but lack real-time control over exposure dynamics. This paper proposes an exposure-aware re-ranking framework operating at the retrieval stage—first decoupling item exposure probability from user interest to enable online, tunable trade-offs between fairness and engagement. Our method explicitly models exposure effects, introduces retrieval-level intervention mechanisms, and implements real-time re-ranking—overcoming key limitations of traditional inverse propensity scoring (IPS) and outcome-preserving calibration (OPC), which are confined to offline settings or post-retrieval ranking layers. A/B testing demonstrates a 25% increase in unique retrieved items, a 40% reduction in head-item dominance, and stable overall user engagement.

Technology Category

• Machine Learning: Learning Preferences or Rankings
• Data Mining & Knowledge Management: Recommender Systems
• Knowledge Representation and Reasoning: Preferences

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Recommender systems operate in closed feedback loops, where user interactions reinforce popularity bias, leading to over-recommendation of already popular items while under-exposing niche or novel content. Existing bias mitigation methods, such as Inverse Propensity Scoring (IPS) and Off- Policy Correction (OPC), primarily operate at the ranking stage or during training, lacking explicit real-time control over exposure dynamics. In this work, we introduce an exposure- aware retrieval scoring approach, which explicitly models item exposure probability and adjusts retrieval-stage ranking at inference time. Unlike prior work, this method decouples exposure effects from engagement likelihood, enabling controlled trade-offs between fairness and engagement in large-scale recommendation platforms. We validate our approach through online A/B experiments in a real-world video recommendation system, demonstrating a 25% increase in uniquely retrieved items and a 40% reduction in the dominance of over-popular content, all while maintaining overall user engagement levels. Our results establish a scalable, deployable solution for mitigating popularity bias at the retrieval stage, offering a new paradigm for bias-aware personalization.

Problem

• Mitigating popularity bias in recommender systems
• Modeling item exposure probability for fair retrieval
• Balancing fairness and engagement in recommendations

Innovation

• Exposure-aware retrieval scoring approach
• Decouples exposure effects from engagement
• Scalable solution for popularity bias