Treating adverse effects of blockbuster bias on beyond-accuracy quality of personalized recommendations

Abstract

Collaborative filtering recommendation algorithms are vulnerable against the popularity bias, including the most popular items repeatedly into the produced ranked lists. However, the research on popularity bias focuses solely on the number of times items are rated rather than the magnitude of the provided ratings when scrutinizing the adverse effects of such bias. This paper introduces a metric describing the blockbuster items that are popular and highly rated simultaneously and investigates the potential biases of collaborative recommendation algorithms towards such items comprehensively. Then, we develop an algorithmic post-processing debiasing approach for potential blockbuster bias in recommendations. Specifically, this method aims to penalize blockbuster items in produced ranked lists by re-sorting items based on the artificial ranking scores, estimated by considering both the blockbuster degree of the items and the generated predictions for them simultaneously. The experiments conducted on three benchmark real-world datasets demonstrate that four prominent collaborative filtering algorithms lead to an undesirable bias in their recommendations towards blockbuster items. The empirical outcomes also indicate that our mitigation method helps treat the adverse effects of the blockbuster bias in terms of beyond-accuracy evaluations such as catalog coverage, diversity, and novelty, with negligible losses in ranking accuracy.