Managing Popularity Bias in Recommender Systems with Personalized Re-ranking Himan Abdollahpouri Robin Burke Bamshad Mobasher [email protected] [email protected] [email protected] Department of Information Science Department of Information Science Web Intelligence Lab, DePaul University of Colorado Boulder University of Colorado Boulder University ABSTRACT Many recommender systems suffer from popularity bias: popular items are recommended frequently while less popular, niche prod- ucts, are recommended rarely or not at all. However, recommending the ignored products in the “long tail” is critical for businesses as they are less likely to be discovered. In this paper, we introduce a personalized diversification re-ranking approach to increase the representation of less popular items in recommendations while main- taining acceptable recommendation accuracy. Our approach is a post-processing step that can be applied to the output of any recom- mender system. We show that our approach is capable of managing popularity bias more effectively, compared with an existing method based on regularization. We also examine both new and existing metrics to measure the coverage of long-tail items in the recommen- Figure 1: The long-tail of item popularity. dation. items are the very popular items, such as blockbuster movies in a ACM Reference Format: Himan Abdollahpouri, Robin Burke, and Bamshad Mobasher. 2019. Manag- movie recommender system, that garner much more viewer attention. ing Popularity Bias in Recommender Systems with Personalized Re-ranking. Similar distributions can be found in other consumer domains. In AAAI Florida Artificial Intelligence Research Society (FLAIRS ’19), May The second vertical line divides the tail of the distribution into two 18–22, 2019, Sarasota, Florida, USA. ACM, New York, NY, USA, 6 pages. parts. We call the first part the long tail: these items are amenable to https://doi.org/10.1145/XXXXXXX collaborative recommendation, even though many algorithms fail to include them in recommendation lists. The second part, the distant INTRODUCTION tail, are items that receive so few ratings that meaningful cross-user comparison of their ratings becomes unreliable. For these cold-start Recommender systems have an important role in e-commerce and items, content-based and hybrid recommendation techniques must information sites, helping users find new items. One obstacle to the be employed. Our work in this paper is concerned with collaborative effectiveness of recommenders is the problem of popularity bias [6]: recommendation and therefore focuses on the long-tail segment. collaborative filtering recommenders typically emphasize popular We present a general and flexible approach for controlling the items (those with more ratings) over other “long-tail” items [16] balance of item exposure in different portions of the item catalog as a that may only be popular among small groups of users. Although post-processing phase for standard recommendation algorithms. Our popular items are often good recommendations, they are also likely work is inspired by [18] where authors introduced a novel probabilis- to be well-known. So delivering only popular items will not enhance tic framework called xQuAD for Web search result diversification new item discovery and will ignore the interests of users with niche which aims to generate search results that explicitly account for tastes. It also may be unfair to the producers of less popular or newer arXiv:1901.07555v4 [cs.IR] 12 Aug 2019 various aspects associated with an under-specified query. We adapt items since they are rated by fewer users. the xQuAD approach to the popularity bias problem. Our approach Figure 1 illustrates the long-tail phenomenon in recommender enables the system designer to tune the system to achieve the desired systems. The y axis represents the number of ratings per item and the trade-off between accuracy and better coverage of long-tail, less x axis shows the product rank. The first vertical line separates the top popular items. 20% of items by popularity – these items cumulatively have many more ratings than the 80% tail items to the right. These “short head” Related Work Permission to make digital or hard copies of part or all of this work for personal or Recommending serendipitous items from the long tail is generally classroom use is granted without fee provided that copies are not made or distributed considered to be a key function of recommendation [5], as these for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. are items that users are less likely to know about. Authors in [7] For all other uses, contact the owner/author(s). showed that 30-40% of Amazon book sales are represented by titles FLAIRS ’19, May 18–22, 2019, Sarasota, Florida, USA that would not normally be found in brick-and-mortar stores. © 2019 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-XXXXXX. Long-tail items are also important for generating a fuller under- https://doi.org/10.1145/XXXXXXX standing of users’ preferences. Systems that use active learning to FLAIRS ’19, May 18–22, 2019, Sarasota, Florida, USA Himan Abdollahpouri, Robin Burke, and Bamshad Mobasher explore each user’s profile will typically need to present more long that adds a personalized bonus to the items that belong to the under- tail items because these are the ones that the user is less likely to represented group (i.e. the long-tail items). The personalization fac- know about, and where user’s preferences are more likely to be tor is determined based on each user’s historical interest in long-tail diverse [17]. items. Finally, long-tail recommendation can also be understood as a social good. A market that suffers from popularity bias will lack METHODOLOGY opportunities to discover more obscure products and will be, by defi- We build on the xQuAD model to control popularity bias in recom- nition, dominated by a few large brands or well-known artists [11]. mendation algorithms. We assume that for a given user u, a ranked Such a market will be more homogeneous and offer fewer opportu- recommendation list R has already been generated by a base recom- nities for innovation and creativity. mendation algorithm. The task of the modified xQuAD method is to The idea of the long-tail of item popularity and its impact on produce a new re-ranked list S (jSj < jRj) that manages popularity recommendation quality has been explored by some researchers bias while still being accurate. [7, 16]. In those works, authors tried to improve the performance of The new list is built iteratively according to the following crite- the recommender system in terms of accuracy and precision, given rion: the long-tail in the ratings. Our work, instead, focuses on reducing P¹vjuº + λP¹v; S 0juº (1) popularity bias and balancing the representation of items across the where P¹vjuº is the likelihood of user u 2 U being interested in popularity distribution. item v 2 V , independent of the items on the list so far as, predicted A regularization-based approach to improving long tail recom- by the base recommender. The second term P¹v; S 0juº denotes the mendations is found in [2]. One limitation with that work is that likelihood of user u being interested in an item v as an item not in this work is restricted to factorization models where the long-tail the currently generated list S. preference can be encoded in terms of the latent factors. In contrast, Intuitively, the first term incorporates ranking accuracy while the a re-ranking approach can be applied to the output of any algo- second term promotes diversity between two different categories of rithm. Another limitation of that work is that it does not account items (i.e. short head and long tail). The parameter λ controls how for user tolerance towards long-tail items: the fact that there may be strongly controlling popularity bias is weighted in general. The item some users only interested in popular items. In our model, we take that scores most highly under the equation 1 is added to the output personalization of long-tail promotion into account as well. list S and the process is repeated until S has achieved the desired And finally, there is substantial research in recommendation di- length. versity, where the goal is to avoid recommending too many similar To achieve more diverse recommendation containing items from items [10, 23, 24]. Personalized diversity is also another related both short head and long tail items, the marginal likelihood P¹v; S 0juº area of research where the amount of diversification is dependent over both item categories long-tail head (Γ) and short head (Γ’) is on the user’s tolerance for diversity [12, 21]. Another similar work computed by: to ours is [20] where authors used a modified version of xQuAD called relevance based xQuAD for intent-oriented diversification of Õ P¹v; S 0juº = P¹v; S 0jdºP¹djuº (2) search results and recommendations. Another work used a similar 2f 0 g approach but for fairness-aware recommendation [14] where xQuAD d Γ;Γ was used to make a fair representation of items from different item Following the approach of [18], we assume that the remaining providers. Our work is different from all these previous diversifica- items are independent of the current contents of S and that the items are independent of each other given the short head and long tail tion approaches in that it is not dependent on the characteristics of 0 items, but rather on the relative popularity of items. categories. Under these assumptions, we can compute P¹v; S jdº in Eq.2 as Ö P¹v; S 0jdº = P¹vjdºP¹S 0jdº = P¹vjdº ¹1 − P¹ijd; Sºº (3) CONTROLLING POPULARITY BIAS i 2S By substituting equation 3 into equation 2, we can obtain xQuAD Õ Ö Result diversification has been studied in the context of information score = ¹1−λºP¹vjuº+λ P¹cjuºP¹vjcº ¹1−P¹ijc; Sºº (4) 0 retrieval, especially for web search engines, which have a similar c 2fΓ;Γ g i 2S goal to find a ranking of documents that together provide a complete where P¹vjdº is equal to 1 if v 2 d and 0 otherwise.