Reproduce “Modeling and Counteracting Exposure Bias in Recommender Systems” on OpenResearch

About this paper

Sami Khenissi, Olfa Nasraoui

This research from the University of Louisville addresses exposure bias in recommender systems, a phenomenon where users primarily interact with already-exposed items, leading to reduced diversity and filter bubbles. The authors propose PEAR-MF, a debiasing strategy that integrates exposure awareness directly into matrix factorization, demonstrating its ability to improve recommendation novelty and diversity over time while maintaining quality.

Key insights

• Exposure bias in recommender systems creates a self-reinforcing cycle where popular items receive more interactions, leading to filter bubbles and reduced diversity over time.
• Modeling user exposure probabilities, particularly with Poisson Factorization, is crucial for understanding and counteracting exposure bias effectively.
• Mitigating exposure bias can significantly improve the long-term novelty and diversity of recommendations without sacrificing overall quality.

Summary by alphaXiv.