The Unfairness of Collaborative Filtering Algorithms’ Bias Towards Blockbuster Items

Abstract

It is known that collaborative filtering recommendation algorithms are usually biased towards some particular items (e.g., popular) in their produced ranked lists. In this study, we evaluate this problem from the perspective of blockbuster items that are both popular and highly-rated items. To this end, we first adopt an efficient method describing blockbuster items, and show that many prominent collaborative filtering algorithms recommend such items too frequently while not giving enough chance to other ones. Furthermore, we evaluate such a blockbuster bias of algorithms from users’ point of view: we comprehensively analyze users’ original propensities in blockbuster items and how such a bias causes the recommendations to deviate from what users desire to receive from the system. To this end, we define three different groups of users based on their interest level in blockbuster items and investigate the impact of this bias on the users in each group. Experimental studies conducted on a real-world benchmark dataset conclude that, for the utilized collaborative filtering algorithms, the recommendations the users receive are highly concentrated on blockbuster items even if users are interested in non-blockbuster ones, showing a significant disparity of observed blockbuster bias.