The Connection Between Popularity Bias, Calibration, and Fairness in Recommendation

Abstract

Recently there has been a growing interest in fairness-aware recommender systems including fairness in providing consistent performance across different users or groups of users. A recommender system could be considered unfair if the recommendations do not fairly represent the tastes of a certain group of users while other groups receive recommendations that are consistent with their preferences. In this paper, we use a metric called miscalibration for measuring how a recommendation algorithm is responsive to users' true preferences and we consider how various algorithms may result in different degrees of miscalibration for different users. In particular, we conjecture that popularity bias which is a wellknown phenomenon in recommendation is one important factor leading to miscalibration in recommendation. Our experimental results using two real-world datasets show that there is a connection between how different user groups are affected by algorithmic popularity bias and their level of interest in popular items. Moreover, we show that the more a group is affected by the algorithmic popularity bias, the more their recommendations are miscalibrated.

Rank-sensitive proportional aggregations in dynamic recommendation scenarios

“…Similar re-formulation can be used for variants of catalogue coverage metrics as well if one-hot representation of items is considered. Studies focused on the popularity bias and calibration phenomena often aim to minimize some form of disproportionality between the proposed recommendations and user profiles (Abdollahpouri et al 2020;Steck 2018). Also, methods dealing with the filter bubbles phenomenon mostly focus on some form of similarity relaxation among recommended items (Lunardi et al 2020), or introduce additional optimization axis less correlated with the estimated relevance of items (Symeonidis et al 2019).…”

“…Several papers focused on an interplay between calibration and other related concepts (Lin et al 2020;Kaya and Bridge 2019;Abdollahpouri et al 2020). Kaya and Bridge (2019) focus on the comparison between calibration-enhancing approaches and intent-aware approaches, which aim to increase the diversity of results.…”

“…Even though the recommendations might had been interesting and relevant at first, the lack of novelty and diversity eventually deteriorates user's experience. Related problems of filter bubbles (Nguyen et al 2014), echo chambers (Ge et al 2020) or popularity bias (Abdollahpouri and Burke 2019;Abdollahpouri et al 2020) also continue to receive considerable attention from the research community.…”

İşbirlikçi Filtreleme Algoritmalarının Çok-Beğenilen Ürünlere Yönelik Yanlılığı

“…Öneri sistemleri, çeşitli alanlardaki birçok uygulamada büyük etkiler oluştursa da, üstesinden gelinmesi zor olan ve öneri etkinliğini bozabilecek birçok yanlılık sorunuyla karşı karşıyadır. Son yıllarda, öneri algoritmalarının sebep olduğu bu tür yanlılıkların nedenleri, öneri kalitesi üzerindeki olumsuz etkileri ve bu etkilerin hafifletilmesi konuları giderek önem kazanmış ve bu alanda yapılan birçok araştırmanın ilgi odağı olmuştur [7]. Şekil 1'de gösterildiği gibi, geri bildirim döngüsünün farklı aşamalarında ortaya çıkabilecek bazı yanlılık türleri tanımlanmıştır.…”

“…İF algoritmaları kullanıcı × ürün derecelendirme matrisi üzerinde eğitildikleri için, içerisinde bulunan tercihlerin karakteristik özelliklerinden etkilenerek ürünler açısından adil olmayan önerilerin üretilmesine neden olabilmektedir [7]. Bu durumun temel nedeni, bu matrislerin doğası gereği ürünler arasında eşit olmayan bir şekilde dağılmış kullanıcı tercihlerini içermesidir.…”