Blockbuster: A New Perspective on Popularity-bias in Recommender Systems

Collaborative filtering algorithms unwittingly produce ranked lists where a few popular items are recommended too frequently while the remaining vast amount of items get not deserved attention, also referred to as the popularity bias problem. Nevertheless, when investigating popularity bias issues in recommendations, the literature commonly estimates the popularity of the items only based on the number of ratings they have received by disregarding their magnitudes. In this study, we evaluate this problem from a different perspective by evaluating the popularity and liking-degree of the items at the same time. To this end, we first develop a method describing the blockbuster items that are both popular and strongly desired by users and propose a helpful metric that measures the potential biases of collaborative filtering algorithms towards such blockbuster items in their produced ranking-based recommendations correctly. The conducted broad set of experiments demonstrate that two popular real-world datasets are highly imbalanced towards blockbuster items. Also, most prominent collaborative algorithms, especially matrix factorization-based SVD and SVD++, lead to an undesirable bias in their recommendations towards blockbuster items.