Incident 168: Collaborative Filtering Prone to Popularity Bias, Resulting in Overrepresentation of Popular Items in the Recommendation Outputs

Description: Collaborative filtering prone to popularity bias, resulting in overrepresentation of popular items in the recommendation outputs.
Alleged: Facebook , LinkedIn , YouTube , Twitter and Netflix developed and deployed an AI system, which harmed Facebook users , LinkedIn users , YouTube users , Twitter Users and Netflix users.

Suggested citation format

Anonymous. (2022-03-01) Incident Number 168. in McGregor, S. (ed.) Artificial Intelligence Incident Database. Responsible AI Collaborative.

Incident Stats

Incident ID
168
Report Count
2
Incident Date
2022-03-01
Editors
Sean McGregor, Khoa Lam

Tools

New ReportNew ReportDiscoverDiscover

Incidents Reports

Introduction

Collaborative filtering (CF) is one of the most traditional but also most powerful concepts for calculating personalized recommendations [22] and is vastly used in the field of multimedia recommender systems (MMRS) [11]. However, one issue of CF-based approaches is that they are prone to popularity bias, which leads to the overrepresentation of popular items in the recommendation lists [2,3]. Recent research has studied popularity bias in domains such as music [15,16] or movies [3] by comparing the recommendation performance for different user groups that differ in their inclination to mainstream multimedia items. However, a comprehensive study of investigating popularity bias on the item and user level across several multimedia domains is still missing (see Section 2).

In the present paper, we therefore build upon these previous works and expand the study of popularity bias to four different domains of MMRS: music (Last.fm), movies (MovieLens), digital books (BookCrossing), and animes (MyAnimeList). Within these domains, we show that users with little interest into popular items tend to have large user profiles and thus, are important consumers and data sources for MMRS. Furthermore, we apply four different CF-based recommendation algorithms (see Section 3) on our four datasets that we each split into three user groups that differ in their inclination to popularity (i.e., LowPop, MedPop, and HighPop). With this, we address two research questions (RQ):

– RQ1: To what extent does an item’s popularity affect this item’s recommendation frequency in MMRS?

– RQ2: To what extent does a user’s inclination to popular items affect the quality of MMRS?

Regarding RQ1, we find that the probability of a multimedia item to be recommended strongly correlates with this items’ popularity. Regarding RQ2, we find that users with less inclination to popularity (LowPop) receive statistically significantly worse multimedia recommendations than users with medium (MedPop) and high (HighPop) inclination to popular items (see Section 4). Our results demonstrate that although users with little interest into popular items tend to have the largest user profiles, they receive the lowest recommendation accuracy. Hence, future research is needed to mitigate popularity bias in MMRS, both on the item and the user level.

Related Work

This section presents research on popularity bias that is related to our work. We split these research outcomes in two groups: (i) work related to recommender systems in general, and (ii) work that focuses on popularity bias mitigation techniques.

Popularity bias in recommender systems. Within the domain of recommender systems, there is an increasing number of works that study the effect of popularity bias. For example, as reported in [8], bias towards popular items can affect the consumption of items that are not popular. This in turn prevents them to become popular in the future at all. That way, a recommender system is prone to ignoring novel items or the items liked by niche users that are typically hidden in the “long-tail” of the available item catalog. Tackling these long-tail items has been recognized by some earlier work, such as [10,20]. This issue is further investigated by [1,2] using the popular movie dataset MovieLens 1M. The authors show that more than 80% of all ratings actually belong to popular items, and based on this, focus on improving the trade-off between the ranking accuracy and coverage of long-tail items. Research conducted in [13] illustrates a comprehensive algorithmic comparison with respect to popularity bias. The authors analyze multimedia datasets such as MovieLens, Netflix, Yahoo!Movies and BookCrossing, and find that recommendation methods only consider a small fraction of the available item spectrum. For instance, they find that KNN-based techniques focus mostly on high-rated items and factorization models lean towards recommending popular items. In our work, we analyze an even larger set of multimedia domains and study popularity bias not only on the item but also on the user level.

Popularity bias mitigation techniques. Typical research on mitigating popularity bias performs a re-ranking step on a larger set of recommended candidate items. The goal of such post-processing approaches is to better expose longtail items in the recommendation list [2,4,6]. Here, for example, [7] proposes to improve the total number of distinct recommended items by defining a target distribution of item exposure and minimizing the discrepancy between exposure and recommendation frequency of each item. In order to find a fair ratio between popular and less popular items, [24] proposes to create a protected group of long-tail items and to ensure that their exposure remains statistically indistinguishable from a given minimum. Beside focusing on post-processing, there are some in-processing attempts in adapting existing recommendation algorithms in a way that the generated recommendations are less biased toward popular items. For example, [5] proposes to use a probabilistic neighborhood selection for KNN methods, or [23] suggests a blind-spot-aware matrix factorization approach that debiases interactions between the recommender system and the user. We believe that the findings of our paper can inform future research on choosing the right mitigation technique for a given setting.

...

Popularity Bias in Collaborative Filtering-Based Multimedia Recommender Systems

If you’re interested in obscure things, there are two reasons why your searches for items and products are likely to be less related to your interests than those of your ‘mainstream’ peers; either you’re a monetization ‘edge case’ whose interests will only be catered to if you’re also in the upper categories of economic purchasing power (for example, products and services related to ‘wealth management’); or the search algorithms that you’re using are leveraging collaborative filtering (CF), which favors the interests of the majority.

Since collaborative filtering is cheaper and more established than other potentially more capable algorithms and frameworks, it’s possible that both these cases apply.

CF-based search results will prioritize items that are perceived to be popular among ‘people like you’, as best the host framework can understand what kind of a consumer you are.

If you’re wary of providing data profiling information to the host system – for instance, not inclined to press the ‘Like’ buttons in Netflix and other video content services – you’re likely to be classified quite generically in your earliest interactions with the system, and the recommendations you receive will reflect the most popular trends.

On a streaming platform, that could mean being recommended whatever shows and movies are currently ‘hot’, such as reality TV and forensic murder documentaries, irrespective of your interest in these. Likewise for book recommendation platforms, which will tend to proffer current and recent best-sellers, apparently arbitrarily.

In theory, even data-circumspect users should eventually get better results from such systems based on the way that they use them and the things that they search for, since most search frameworks give users limited ability to edit their usage history.

Any Color You Like, so Long as It’s Black

However, according to a new study from Austria, the ascendancy of collaborative filtering over content-based filtering (which seeks to define relationships between products instead of just taking aggregate popularity into account), and other alternative approaches, inclines search systems towards long term popularity bias, where obviously popular results are pushed towards end users that are unlikely to be enthused by them.

The paper finds that users who are uninterested in popular items receive ‘significantly worse’ recommendations than users with medium or high interest in popularity, and (perhaps tautologically) that popular items are recommended more frequently than unpopular items. The researchers also conclude that users with low interest in popular items tend to have larger user profiles that could potentially improve recommender systems – if only the systems could kick their addiction to ‘herd’ metrics.

The paper is titled Popularity Bias in Collaborative Filtering-Based Multimedia Recommender Systems, and comes from researchers at now-Center GmbH in Graz, and the Graz University of Technology.

Domains Covered

Building on prior works that studied individual sectors (such as book recommendations), the new paper examines four domains: digital books (via the BookCrossing dataset); movies (via MovieLens); music (via Last.fm); and animes (via MyAnimeList).

The study applied four popular multimedia recommender systems (MMRS) collaborative filtering algorithms against datasets split into three user groups, according to their inclination to be receptive to ‘popular’ results: LowPop, MedPop, and HighPop. The user groups were filtered down to 1000 equal size groups, based on least, average, and most likely to favor ‘popular’ results.

Commenting on the results, the authors state:

‘[We] find that the probability of a multimedia item to be recommended strongly correlates with this items’ popularity [and] that users with less inclination to popularity (LowPop) receive statistically significantly worse multimedia recommendations than users with medium (MedPop) and high (HighPop) inclination to popular items…

‘Our results demonstrate that although users with little interest into popular items tend to have the largest user profiles, they receive the lowest recommendation accuracy. Hence, future research is needed to mitigate popularity bias in MMRS, both on the item and the user level.’

Among the algorithms evaluated were two K-Nearest Neighbors (KNN) variants, UserKNN and UserKNNAvg. The first of these does not generate an average rating for the target user and item. A non-negative matrix factorization variant (NMF) was also tested, along with a CoClustering algorithm.

The evaluation protocol considered the recommendation task as a prediction challenge, measured by the researchers in terms of mean absolute error (MAE), against a five-fold cross validation protocol that exceeds the usual 80/20 split between trained and test data.

The results indicate a near-guarantee of popularity bias under collaborative filtering. The question, arguably, is whether this is perceived as a problem by the multi-billion dollar companies currently incorporating CF into their search algorithms.

The ‘Easy’ Way Out

Though collaborative filtering is increasingly used as only one plank of a broader search algorithm strategy, it has a strong stake in the search sector, and its logic and potential profitability is attractively easy to understand.

In itself, CF essentially offloads the task of evaluating content value to end users, and uses their uptake of the content as an index of its value and potential attractiveness to other customers. By analogy, it’s essentially a map of ‘water cooler buzz’.

Content-based filtering (CBF) is more difficult, but could potentially provide more relevant results. In the computer vision sector, an increasing amount of research is currently being expended on categorizing video content and attempting to derive domains, features, and high level concepts through analysis of audio and video in movie and TV output.

However, this is a relatively nascent pursuit, and bound up in the current, more general struggle to quantify, isolate and exploit high level concepts and features in domain knowledge.

Who Uses Collaborative Filtering?

At the time of writing, Netflix’s oft-criticized recommendation engine remains fixated on various collaborative filtering approaches, applying a variety of adjunct technologies in ongoing attempts to generate more user-relevant recommendations.

Amazon’s search engine evolved from its early adoption of user-based collaborative filtering to an item-item collaborative filtering method, which places greater emphasis on the customer’s purchase history. Naturally, this can lead to different types of inaccuracy, such as filter bubbles, or over-emphasis on sparse data. In the latter case, if an infrequent Amazon customer makes an ‘unusual’ purchase, such as a set of operettas for an opera-loving friend, there may not be adequate alternative purchases that reflect the customer’s own preferences to stop this purchase from becoming an influence on their own recommendations.

Collaborative filtering is also extensively used by Facebook, in concert with other approaches, and also by LinkedIn, YouTube, and Twitter.

Why AI Isn’t Providing Better Product Recommendations