Exposure Concentration and Bias Accumulation in Video Recommendation Platforms

Abstract

Personalized recommendation systems aim to improve user experience by tailoring content delivery. However, long-term operation may gradually introduce systematic biases that disadvantage specific user groups or content providers. This research examines interaction records from a video streaming platform serving approximately 4.6 million active users. Exposure rates, click-through patterns, and content diversity metrics were tracked over 18 months. Analysis reveals that recommendation concentration increased steadily, with top 5% of creators receiving over 54% of total impressions by the end of the study period. Several re-ranking and diversity-aware strategies were tested in an online sandbox environment. While short-term engagement declined slightly, content visibility became more evenly distributed.