Popularity Bias in Hotel Recommendation: Exploring the Impact of Calibration on User Preferences and diversity

Beld, van den, Esmee

Popularity Bias in Hotel Recommendation: Exploring the Impact of Calibration on User Preferences and diversity

Files

!Beld, van den, E. s-103782-BSc-Thesis-2022.pdf (1.07 MB)

Authors

Beld, van den, Esmee

Issue Date

2022-06-28

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/15985

Abstract

Recommender systems are algorithms used to find items to recommend to a user based on their interest. These systems are known to suffer from popularity bias, which is the tendency of the system to unequally recommend a set of popular items more, and to recommend the items outside of this set, less, even though these items would be preferred by the user. By popularity bias, mismatches appear between the interest of a user and the recommendations that they receive. To decrease the impact of popularity bias in recommender systems and with that, lower the amount of mismatches, the application of calibration is suggested. Calibration in recommender systems is a re-ranking method that matches the distribution of the recommendations to the distribution in the user’s historical behaviour over a chosen category. To maintain a healthy and sustainable consumption pattern, which is necessary for the success of a platform, a more diverse list of recommendations is important. To see if the diversity changes after the application of calibration, diversity measures are applied to the given recommendations. Popularity bias is found to appear in hotel recommendation data, where users with different interests in item popularity are unequally impacted. Differing amounts of mismatches between the user’s interest and their received recommendations appear between the separate user groups. Users with an interest in non-popular items are impacted more severely than those with an interest in popular items. After applying calibration, mismatches between users interest in a certain hotel type and the recommendations that the users got, are decreased. On the old and new list of recommendations, the diversity metrics are applied and there is found that diversity is increased after applying calibration.