Detection and Mitigation of Popularity Bias in E-Commerce Recommender Systems

Popularity bias is a persistent obstacle in e-commerce recommender systems. Under heavy-tailed implicit feedback and exposure-driven feedback loops, ranking models tend to over-recommend already visible products and to suppress long-tail items that may be relevant to users with heterogeneous preferences. Existing mitigation strategies usually intervene at one point in the pipeline, for example by modifying the training loss or by re-ranking final lists, and therefore often face an unstable accuracy–fairness trade-off. This paper proposes a unified framework for detecting and mitigating popularity bias in e-commerce recommendation. The framework couples three components: multi-level diagnostics for recommendation concentration, interaction-aware training constraints for score calibration under uneven exposure, and exposure-balanced re-ranking for serving-time control. The key contribution is not a single isolated module, but the structural coupling between upstream representation calibration and downstream exposure allocation. Offline evaluations on RetailRocket, Amazon Electronics 5-core, and TAOBAO-MM show that the unified framework lowers Average Recommendation Popularity and Gini concentration while increasing catalogue coverage, with only negligible loss in nDCG and Recall. Additional ablation, budget, user-segment, and scalability analyses clarify when the method improves long-tail exposure without injecting irrelevant tail items.