Managing popularity bias in multi-criteria recommender systems

Multi-criteria recommendation systems aim to provide more personalized and comprehensive recommendations by allowing users to evaluate products or services ba- sed on multiple criteria. However, one of the fundamental problems encountered in these systems is popularity bias. Popularity bias occurs when collective tendencies become do- minant over individual preferences; this situation leads to over-representation of popular content in recommendation lists and the neglect of less-known but suitable alternatives. In this context, this thesis proposes two innovative approaches, RelPref and PriRelPref, aimed at reducing popularity bias, in addition to examining existing methods. These met- hods model user-item interactions more sensitively and balancedly, taking into account users' criterion priorities. While RelPref dynamically evaluates user scores for each crite- rion, PriRelPref more effectively reduces bias by incorporating user-specific criterion pri- oritization. Thus, harmony is achieved between the criteria that users value most and the criteria where items perform best, providing more balanced, diverse, and personalized recommendations.The proposed methods have been comprehensively tested on three dif- ferent datasets (YM10, YM20, and TA) representing different user rating behaviors. Multi-dimensional metrics (such as Gini index for long-tail item visibility and diversity) have shown that RelPref and PriRelPref significantly improve both user satisfaction and system-level qualities such as discoverability and fair content distribution. Particularly, PriRelPref's criterion prioritization feature provides a more effective solution in managing popularity bias. In conclusion, this study is one of the first studies to systematically address popu- larity bias in multi-criteria recommendation systems, filling an important gap in the lite- rature and establishing a solid theoretical and methodological foundation for future rese- arch.