PI-adaptDiv: an adaptive algorithm to prevent and escape online filter bubbles

Recommender systems are essential for managing the abundance of online content and enhancing user engagement through personalized suggestions. However, their reliance on personalization risks reinforcing ”filter bubbles,” a phenomenon where users repeatedly encounter content aligning closely with their existing preferences. This cycle can lead to a reduction in diversity, limiting exposure to novel ideas and exacerbating polarization and misinformation. Traditional recommender system approaches to mitigate filter bubbles typically emphasize recommended diversity, balancing the variety of items shown to users while maintaining relevance. Yet, a significant gap exists between recommended diversity (what users see) and consumed diversity (what users choose). Even when recommendations are diversified, users exhibiting strong selective exposure may continue to select content aligned with their existing preferences, thereby perpetuating filter bubbles. To address this issue, we propose PI-adaptDiv, an algorithm designed to increase consumed diversity directly rather than merely diversifying recommendations. PI-adaptDiv employs a Proportional-Integral (PI) controller to dynamically adjust recommendation diversity based on user interactions, prioritizing items that enhance consumed diversity while maintaining a high likelihood of user selection. Evaluations using offline simulations and an online user experiment on a video recommendation platform demonstrate PI-adaptDiv’s effectiveness in significantly enhancing consumed