Detecting Hidden Bias in AI Recommendation Systems

This paper introduces a framework to evaluate representation bias within latent factor recommendation (LFR) models, focusing on how user and item embeddings may encode implicit associations with sensitive attributes like gender. Unlike prior research that centers on performance metrics or exposure bias, this work examines attribute association bias and demonstrates its measurement through an industry case study in podcast recommendations. The goal is to help practitioners audit, interpret, and mitigate bias propagation across multi-stage recommender pipelines, promoting greater fairness and transparency in AI systems.