Towards Fair Multi-Stakeholder Recommender Systems

Francois Buet-Golfouse, Islam Utyagulov

Abstract: Since recommender systems (“RSs”) are used in multiple domains and applications, issues of possible biases and discrimination have become paramount and present a technical challenge. Indeed, fairness in and for RSs is specific as it does not only concern protected attributes but also encompasses notions of user representativeness and item diversity. In short, RSs require multi-stakeholder fairness. However, RSs are generally built on large (and possibly very sparse) datasets, thus precluding the use of very complex debiasing techniques. Our approach introduces a fairness functional that minimises the loss disparity across groups and avoids a post-processing step. This enables us to adapt most algorithms underlying recommender systems, such as factorisation machine and its many generalisations and debias them. We present an example of such a functional and show that its properties are ideally suited to the case of multi-stakeholder fair RS. Finally, we demonstrate that our approach works well in practice on benchmark datasets and that partial debiasing is essential, as full debiasing may lead to poor generalisation.