Developing novel methods to mitigate the adverse effects of the popularity bias problem in recommender systems


Bilge A., Yalçin E.(Executive)

TUBITAK Project, 2022 - 2024

  • Project Type: TUBITAK Project
  • Begin Date: June 2022
  • End Date: June 2024

Project Abstract

Producing personalized recommendations in recommender systems is usually performed using collaborative filtering (CF) algorithms due to their success and efficiency. These algorithms are based on the assumption that users with similar propensities in the past will have similar preferences in the future, and they usually focus on producing ranked lists including items that users have not yet experienced.

CF algorithms have some known weaknesses. The first is that users of such systems might think that their privacy is violated due to indicating personal preferences and, therefore, either providing false information to those systems or, even worse, completely denying to use them. Privacy-preserving CF (PPCF) algorithms developed to address users' privacy concerns provide reliable recommendations on masked data using data perturbation, substitution, and manipulation techniques. Another weakness of the CF algorithms is that they produce recommendation lists where popular items appear too frequently, while unpopular ones do not receive the deserved attention, even if they are of reasonable interest. This issue usually originates from the imbalances in the rating distribution in the user-item matrix where CF algorithms are trained.

Considering that PPCF algorithms focus on substituting and perturbating original user profiles with fake ratings, it is expected that they will lead to different popularity bias results compared to CF algorithms. Therefore, the project's first aim is to analyze the popularity bias performance of the PPCF algorithms. In addition, the performance of the popularity-debiasing methods developed for CF algorithms will be investigated in the context of PPCF algorithms. Also, it aims to develop a novel popularity-debiasing method, which is compatible with both CF and PPCF algorithms and will enable the production of more qualified recommendation lists than existing methods in terms of beyond-accuracy perspectives. This method will adopt the data filling strategy used for data privacy in the PPCF algorithms and focus on increasing the rating numbers of unpopular items without disturbing the general structure of the profiles. Thus, it will ensure that algorithms are trained on more balanced profiles in terms of rating distribution, and unpopular items will be featured in the recommendation lists.

Recommender systems should ensure that items recommended to users satisfy them to a maximum extent. Therefore, it is required to carefully analyze user profiles and suggest items related to their past experiences. For that reason, the actual propensities of the users on popular items will be analyzed by considering their different characteristics, and it will be examined how popularity bias in the recommendations imposed by the CF and PPCF algorithms affects users/groups with different characteristics. In addition, it is aimed to develop a novel user-centric popularity-debiasing approach compatible with both CF and PPCF algorithms, which considers the actual popularity tendencies of users. This approach will be based on the re-ranking ranked lists strategy and will be optimized based on the actual popularity interests of users and the popularity of the items. Studies that are to be performed during the proposed project will be novel due to completing a critical shortage in the literature and hence have a high potential of publishing in indexed journals. The produced scientific publications will contribute to the international recognition of the project coordinator in this field and provide the opportunity to conduct post-doctoral research abroad. Also, such solutions will be directly applicable to sectors such as tourism and e-commerce. A principal investigator, a researcher, and two graduate students will be employed within the project. Those two students will complete their dissertations based on their studies during the project.