The Unfairness of Popularity Bias in Book Recommendation

Recent studies have shown that recommendation systems commonly suffer from popularity bias. Popularity bias refers to the problem popular items (i.e., frequently rated items) are recommended while less rarely or not at all. Researchers adopted two approaches examining bias: (i) users’ perspective, by analyzing how far a system deviates user’s expectations in receiving items, and (ii) amount of exposure long-tail receive, measured overall catalog coverage novelty. In this paper, we examine first point view book domain, although findings may be applied other domains as well. To end, analyze well-known Book-Crossing dataset define three user groups based on their tendency towards Niche, Diverse, Bestseller-focused). Further, evaluate performance nine state-of-the-art algorithms baselines Random, MostPop) both accuracy (e.g., NDCG, Precision, Recall) perspectives. Our results indicate most fail meet with Niche Diverse tastes despite having larger profile size. Conversely, Bestseller-focused users more likely receive high-quality recommendations, terms fairness personalization. Furthermore, our study shows tradeoff between personalization unfairness for belonging Bestseller groups, is, high capability Finally, across models, show WMF VAECF can provide higher quality when considering