Understanding Attribute Association Bias in Recommender Systems

This paper introduces a practical evaluation framework for detecting attribute association bias (AAB) in latent factor recommendation systems — a subtle but critical form of representation bias where sensitive attributes (like gender) become entangled in model embeddings. By adapting methods from NLP bias research, the authors propose four evaluation strategies—bias vector creation, AAB metrics, classification explanations, and latent space visualization—to help practitioners quantify and interpret bias in recommendation outputs. Tested on a real-world podcast recommendation model, the framework exposes significant user gender bias even after mitigation attempts, underscoring the need for systematic AAB audits in AI-driven personalization systems.