Akademik Veri Yönetim Sistemi
Journal of Intelligent Information Systems, 2026 (SCI-Expanded, Scopus)
Ensuring equitable treatment of different user groups in recommender systems is a key challenge, and the issue of fairness has been widely explored in the literature. However, understanding fairness within a robust feedback loop, as it occurs in real-world settings, remains elusive. This study examines the interplay between popularity bias, calibration, accuracy, and beyond-accuracy performance of recommender systems using a novel dynamic feedback loop framework. The framework models iterative interactions between recommendation algorithms and user profiles, enabling the analysis of calibration, accuracy, and beyond-accuracy measures across user groups with varying preferences on popular items, i.e., Popular-, Diverse-, and Niche-focused. Empirical evaluations conducted on two benchmark datasets using three collaborative filtering algorithms reveal distinct disparities in how feedback loops affect different user groups. Niche-focused users, despite being the most active and information-rich, experience the steepest deterioration in system alignment over time, losing much of their initial calibration, long-tail exposure, and diversity advantages, along with proportional declines in accuracy. These results show that feedback dynamics progressively misalign the system with its most valuable users, making them the most disadvantaged over time. Popular-focused users remain most aligned with algorithmic tendencies, achieving steady accuracy gains but remaining confined to narrow, popularity-driven content with little to no long-tail exposure. Meanwhile, Diverse-focused users, initially balanced between popular and niche preferences, undergo gradual calibration drift and consistent reductions in diversity and long-tail representation, gradually converging toward recommendation patterns similar to Popular-focused users. Overall, the results demonstrate that feedback loops magnify structural inequalities, reinforcing popularity bias while reducing diversity and personalization across all user groups.