Search for: All records

Recommender systems have a variety of stakeholders. Applying concepts of fairness in such systems requires attention to stakeholders’ complex and often-conflicting needs. Since fairness is socially constructed, there are numerous definitions, both in the social science and machine learning literatures. Still, it is rare for machine learning researchers to develop their metrics in close consideration of their social context. More often, standard definitions are adopted and assumed to be applicable across contexts and stakeholders. Our research starts with a recommendation context and then seeks to understand the breadth of the fairness considerations of associated stakeholders. In this paper, we report on the results of a semi-structured interview study with 23 employees who work for the Kiva microlending platform. We characterize the many different ways in which they enact and strive toward fairness for microlending recommendations in their own work, uncover the ways in which these different enactments of fairness are in tension with each other, and identify how stakeholders are differentially prioritized. Finally, we reflect on the implications of this study for future research and for the design of multistakeholder recommender systems. 

While substantial advances have been made in recommender systems -- both in general and for news -- using datasets, offline analyses, and one-shot experiments, longitudinal studies of real users remain the gold standard, and the only way to effectively measure the impact of recommender system designs (algorithmic and otherwise) on long-term user experience and behavior. While such infrastructure exists for studies within some individual organizations, the extensive cost and effort to build the systems, content streams, and user base make it prohibitive for most researchers to conduct such studies. We propose to develop shared research infrastructure for the research community, and have received funding to gather community input on requirements, resources, and research goals for such an infrastructure. If the full infrastructure proposal is funded, it would result in recruiting a community of thousands of users who agree to use a news delivery application within which various researchers would be install and conduct experiments. In this short paper we outline what we have heard and learned so far and present a set of questions to be directed to INRA attendees to gather their feedback at the workshop. 

Recommender systems are poised at the interface between stakeholders: for example, job applicants and employers in the case of recommendations of employment listings, or artists and listeners in the case of music recommendation. In such multisided platforms, recommender systems play a key role in enabling discovery of products and information at large scales. However, as they have become more and more pervasive in society, the equitable distribution of their benefits and harms have been increasingly under scrutiny, as is the case with machine learning generally. While recommender systems can exhibit many of the biases encountered in other machine learning settings, the intersection of personalization and multisidedness makes the question of fairness in recommender systems manifest itself quite differently. In this article, we discuss recent work in the area of multisided fairness in recommendation, starting with a brief introduction to core ideas in algorithmic fairness and multistakeholder recommendation. We describe techniques for measuring fairness and algorithmic approaches for enhancing fairness in recommendation outputs. We also discuss feedback and popularity effects that can lead to unfair recommendation outcomes. Finally, we introduce several promising directions for future research in this area.

 

Recommendation and ranking systems are known to suffer from popularity bias; the tendency of the algorithm to favor a few popular items while under-representing the majority of other items. Prior research has examined various approaches for mitigating popularity bias and enhancing the recommendation of long-tail, less popular, items. The effectiveness of these approaches is often assessed using different metrics to evaluate the extent to which over-concentration on popular items is reduced. However, not much attention has been given to the user-centered evaluation of this bias; how different users with different levels of interest towards popular items are affected by such algorithms. In this paper, we show the limitations of the existing metrics to evaluate popularity bias mitigation when we want to assess these algorithms from the users’ perspective and we propose a new metric that can address these limitations. In addition, we present an effective approach that mitigates popularity bias from the user-centered point of view. Finally, we investigate several state-of-the-art approaches proposed in recent years to mitigate popularity bias and evaluate their performances using the existing metrics and also from the users’ perspective. Our experimental results using two publicly-available datasets show that existing popularity bias mitigation techniques ignore the users’ tolerance towards popular items. Our proposed user-centered method can tackle popularity bias effectively for different users while also improving the existing metrics. 

Though recommender systems are defined by personalization, recent work has shown the importance of additional, beyond-accuracy objectives, such as fairness. Because users often expect their recommendations to be purely personalized, these new algorithmic objectives must be communicated transparently in a fairness-aware recommender system. While explanation has a long history in recommender systems research, there has been little work that attempts to explain systems that use a fairness objective. Even though the previous work in other branches of AI has explored the use of explanations as a tool to increase fairness, this work has not been focused on recommendation. Here, we consider user perspectives of fairness-aware recommender systems and techniques for enhancing their transparency. We describe the results of an exploratory interview study that investigates user perceptions of fairness, recommender systems, and fairness-aware objectives. We propose three features – informed by the needs of our participants – that could improve user understanding of and trust in fairness-aware recommender systems. 

It is well known that explicit user ratings in recommender systems are biased toward high ratings and that users differ significantly in their usage of the rating scale. Implementers usually compensate for these issues through rating normalization or the inclusion of a user bias term in factorization models. However, these methods adjust only for the central tendency of users’ distributions. In this work, we demonstrate that a lack of flatness in rating distributions is negatively correlated with recommendation performance. We propose a rating transformation model that compensates for skew in the rating distribution as well as its central tendency by converting ratings into percentile values as a pre-processing step before recommendation generation. This transformation flattens the rating distribution, better compensates for differences in rating distributions, and improves recommendation performance. We also show that a smoothed version of this transformation can yield more intuitive results for users with very narrow rating distributions. A comprehensive set of experiments, with state-of-the-art recommendation algorithms in four real-world datasets, show improved ranking performance for these percentile transformations. 

Comparative experimentation is important for studying reproducibility in recommender systems. This is particularly true in areas without well-established methodologies, such as fairness-aware recommendation. In this paper, we describe fairness-aware enhancements to our recommender systems experimentation tool librec-auto. These enhancements include metrics for various classes of fairness definitions, extension of the experimental model to support result re-ranking and a library of associated re-ranking algorithms, and additional support for experiment automation and reporting. The associated demo will help attendees move quickly to configuring and running their own experiments with librec-auto.