Mitigating Popularity Bias in Online Recommendation Systems under Skewed Data Distributions

Abstract

Online recommendation systems are strongly influenced by uneven user activity and content popularity. A small fraction of users and items often dominates training data, leading to biased prediction results. This paper reports observations from an internal recommendation platform serving multimedia content. Analysis shows that less than 7% of items account for more than half of user interactions. Several reweighting and sampling strategies were evaluated to mitigate this effect. Results indicate that moderate performance gains can be achieved without increasing system complexity, but long-tail recommendation remains difficult under highly skewed data distributions.